1.40.0 (Pending)

Incompatible behavior changes

Changes that are expected to cause an incompatibility if applicable; deployment changes are likely required

build: Bumped the hermetic LLVM/Clang toolchain from 18 to 22. This upgrades the default compiler and may surface new warnings or diagnostics in downstream builds that pin to the Envoy toolchain.
build: Removed the obsolete Clang Bazel configuration. Clang with libc++ is now the default toolchain and requires no configuration flag. GCC with libstdc++ remains available via --config=gcc. Other compiler/standard-library combinations require a user-provided toolchain. The --@envoy//bazel:libc++ and --@envoy//bazel:libstdc++ flags and the force_libcpp define have been removed; the standard library is now derived from the compiler. The //bazel:force_libcpp label is retained as an alias for //bazel:libc++_enabled. Also, removed the non-functional, obsolete clang-local Bazel configuration. Downstreams that require a host LLVM toolchain should use the bzlmod example in bazel/tests/codeql; its LLVM version is automatically detected from clang --version.
build: The --define wasm=<engine> and --define engine=<engine> build flags for selecting the WebAssembly runtime have been replaced by the first-class Bazel build setting --@proxy-wasm-cpp-host//bazel:engine=<engine>. Accepted values are v8 (default), wamr (interpreter mode), wamr-interp, wamr-jit, wasmtime, null, disabled, and multi. The old --define flags are no longer honoured and will raise a failure. Users and CI scripts must update their invocations.
dynamic_modules: The Rust dynamic-module SDK’s HttpFilter hooks now take &self instead of &mut self. Existing filters must update their hook signatures and hold any mutable per-stream state behind interior mutability (Cell/RefCell), and their Drop must not panic. This closes a use-after-free: a filter hook that triggered a synchronous teardown of the filter chain (for example recreate_stream) could return to a freed in-module filter. The filter is now reference-counted for the duration of each hook, which requires shared borrows, since Envoy can re-enter the filter synchronously while a hook is still on the stack and two aliasing &mut self would be undefined behavior.
dynamic_modules: dynamic modules: aligned IP source and destination address attributes with CEL by including their ports, preserved non-IP address strings, and made missing connection IDs and addresses unavailable.
dynamic_modules: dynamic modules: fixed xds.virtual_host_name to return the selected virtual host name instead of the virtual cluster name. Added xds.virtual_cluster_name for the virtual cluster name and updated the deprecated access logger callback to use it.
local_ratelimit: Fixed a bug in the local rate limiter where an exhausted shadow_mode: true descriptor would short-circuit descriptor evaluation and prevent the remaining enforced descriptors and the default token bucket from being consumed. Requests that were previously allowed by this bypass may now be rate limited. This behavior can be reverted by setting the runtime guard envoy.reloadable_features.local_ratelimit_shadow_mode_no_short_circuit to false.
mcp_router: Added header_forwarding to McpBackend, in both the MCP router filter and the mcp_multicluster cluster type, giving per-backend control over which downstream request headers are forwarded upstream. The default has also changed: previously all downstream request headers were forwarded to every backend except a small hardcoded skip-list; now, unless a backend explicitly configures header_forwarding, no downstream-controlled headers are forwarded beyond those the router itself synthesizes – in particular, the client’s authorization header is no longer forwarded by default. This matters because a single mcp_router can aggregate backends of mixed trust, and MCP requires audience-bound tokens rather than implicit passthrough. Deployments relying on the previous forward-everything behavior can restore it per backend by setting forward_all: true on header_forwarding.
vhds: The on_demand filter, when performing on-demand VHDS, will no longer recreate the stream after a route configuration update successfully resolves the virtual host. Instead it refreshes the route configuration snapshot and continues decoding the existing stream, so filters appearing before the on_demand filter are no longer invoked twice. This mirrors the existing on-demand CDS behavior gated by envoy.reloadable_features.on_demand_cluster_no_recreate_stream. This behavior can be temporarily reverted by setting the runtime guard envoy.reloadable_features.on_demand_vhds_no_recreate_stream to false.

Minor behavior changes

Changes that may cause incompatibilities for some users, but should not for most

access_log: Integer-valued access log substitution commands are now rendered as exact integers in JSON access logs rather than as shortest-round-trip doubles. Commands whose value is naturally an integer – such as %BYTES_SENT%, %DURATION% and %COMMON_DURATION% – previously went through a double, which is serialized in whichever of the plain and the exponent form is shorter. Only the values that took the exponent form change: those are round numbers with enough trailing zeros, the smallest being 100000, which was emitted as 1e+05 and is now emitted as 100000. Values such as 123456, 1500000 and 86400000 were already emitted in full and are unchanged, as are text (non-JSON) access logs. Consumers that parse JSON access logs with a parser accepting either form see no difference, but a consumer relying on the exponent form needs updating.
access_log: The %COALESCE()% access log operator now returns the first result that is present, including a result that is present but empty. Previously an operator producing an empty value was treated as if it had produced no value at all, and the next operator in the list was evaluated. This change can be reverted by setting the runtime guard envoy.reloadable_features.coalesce_formatter_accept_empty_values to false.
access_log: The length limit Z of the metadata command operators (%DYNAMIC_METADATA()%, %CLUSTER_METADATA()%, %UPSTREAM_METADATA()% and the %METADATA()% formatter extension) now only truncates string values in typed output such as JSON access logs. Previously any non-structured value was rendered as JSON and, when the limit applied, emitted as a truncated string, so a numeric value of 1234 with a limit of 2 was logged as the string "12". Numbers and booleans now keep their type and are logged in full, matching structs and lists, which were already never truncated. Text access logs are unchanged: the rendered value is still truncated to Z characters. This behavioral change can be reverted by setting the runtime guard envoy.reloadable_features.metadata_formatter_only_truncate_string to false.
admin: A non-graceful drain (/drain_listeners without graceful) now starts a drain sequence and notifies the connections of the covered listeners that a drain has begun, in addition to stopping the listeners. Previously nothing was drained: the listeners simply stopped accepting, and the connections they already owned were never told, so no connection-level drain logic ran for them. The drain honors the configured --drain-strategy, as a graceful drain does; graceful only controls whether the listeners keep accepting for a drain period before they are stopped. As a result, skip_exit is now accepted without graceful (it was rejected with a 400 before) and means “drain the connections, but never stop the listeners”, which is what graceful&skip_exit already did. This change can be temporarily reverted by setting the runtime guard envoy.reloadable_features.non_graceful_drain_notifies_connections to false.
build: The google-vrp Docker image variant is no longer published. Envoy continues to participate in Google’s Patch Reward Program (PRP), which does not require a Docker execution environment.
config: The server’s default protobuf message validation visitor, returned by ServerFactoryContext::messageValidationVisitor(), now switches from the static to the dynamic visitor as soon as the bootstrap resources have been loaded, rather than when the main dispatch loop starts.
drain: Connection drain-close decisions (HTTP connection manager, TCP proxy, Mongo proxy, Redis proxy, Thrift proxy, generic proxy and the drain-aware HTTP connection manager) are now derived from a drain event that is pushed to each connection when the drain sequence starts on the main thread, rather than from polling the listener’s DrainDecision on every response.

This also fixes a bug where the Mongo, Redis, Thrift and generic proxies did not honor inbound-only drain-close decisions: they always asked whether both inbound and outbound connections were draining, so /drain_listeners?graceful&inboundonly left their connections open even on an inbound listener. Because the new drain event is only delivered to the listeners covered by the drain, these proxies now drain-close in that case.

This behavioral change can be temporarily reverted by setting the runtime guard envoy.reloadable_features.use_connection_event_drain to false. The guard is read once per connection when the network filter is created, so changing it affects new connections only.
drain: The connections of a draining filter chain (an in-place listener filter chain update, or a filter chain removal) are now notified with the configured drain strategy (--drain-strategy) instead of always being notified with immediate. With the default gradual strategy these connections are now drain-closed by ramping up over the drain window rather than at the first opportunity. This behavioral change can be reverted by setting the runtime guard envoy.reloadable_features.filter_chain_drain_uses_configured_strategy to false.
dynamic_forward_proxy: DNS cache statistics (dns_cache.*) are now always created under the server-wide stats scope instead of a caller-derived scope captured when the manager singleton was first instantiated. As a result they are now matched against the global stats matcher rather than a per-listener stats matcher. Statistic names are unchanged, so this only affects configurations where the listener that first created a DNS cache declared a per-listener stats matcher.
dynamic_modules: Dynamic module load failures now emit an error log on the dynamic_modules logger with the form Unable to load dynamic module <module> <reason>. The log is emitted from the module loader for every extension type, including extension points that have no factory context. The formatter and health checker extensions now pass the server factory context when loading a module by name, so a load failure increments the shared dynamic_modules.module_load_error counter tagged with the configured instance name.
dynamic_modules: Dynamic modules are now loaded with RTLD_NOW so that every referenced symbol is resolved at load time instead of lazily on first use. A module that references a symbol the main program does not provide now fails to load rather than crashing later when the symbol is first reached. This change can be reverted by setting the runtime guard envoy.reloadable_features.dynamic_modules_rtld_now to false, which restores the previous RTLD_LAZY behavior.
dynamic_modules: The Rust dynamic-module SDK’s mock feature now uses mockall 0.15 instead of 0.13. Modules whose tests pass their own mockall types, for example a Sequence, to the SDK’s mock types must update their mockall dependency to 0.15.
dynamic_modules: The dynamic modules tracer now returns the real sampling decision from exportedSpan instead of always reporting the span as exported. When a span sampling decision is set to false through setSampled, the span is now skipped at finalization, so module finalize callbacks such as envoy_dynamic_module_on_tracer_span_set_tag no longer run for it. A span still defaults to exported until setSampled reports otherwise.
dynamic_modules: envoy_dynamic_module_on_http_filter_destroy now runs after the HTTP stream is gone, so the callbacks that need it, for example the ones reading the headers, the stream info or the buffered bodies, are no-ops. Modules that did end of stream bookkeeping from the destroy hook should do it from envoy_dynamic_module_on_http_filter_stream_complete instead.
ext_proc: Changed the stats scope that is used to create the gRPC client of the external processing (ext_proc) filter. Previously the filter’s own scope was used, so the gRPC client stats gained an unexpected extra prefix, cluster.<cluster_name>. for an upstream filter. The server scope is now used, so the Google gRPC client stats are emitted with the expected grpc.<google_grpc_stat_prefix>. prefix.
ext_proc: The external processing filter now logs the target URI as the destination when using the google_grpc service. Previously, only the cluster name of the envoy_grpc service was logged.
ext_proc: ext_proc: Avoid sending empty data frames with end_stream=false to the external processing server. The Envoy filter manager sometimes generates these frames, which Envoy consumes internally without forwarding to the backend. Skipping these frames during external processing as well. This behavior can be reverted by setting the runtime guard envoy.reloadable_features.ext_proc_not_send_empty_data_with_false_eos to false.
gcp_authn: The filter’s statistics are now emitted in the http.<stat_prefix>.gcp_authn. namespace instead of directly in the HTTP connection manager’s http.<stat_prefix>. namespace, matching the convention used by the other HTTP filters. For example http.ingress.retrieve_audience_failed is now http.ingress.gcp_authn.retrieve_audience_failed. Dashboards and alerts referring to the old names need to be updated.
geoip: The MaxMind geolocation provider now loads each geolocation database file once and shares it between every provider configured with that file, rather than loading a separate copy per provider. As part of this, the database file level statistics db_reload_success, db_reload_error and db_build_epoch are no longer emitted under the listener derived <stat_prefix>.maxmind. namespace. They are now emitted as maxmind.<db_type>.<db_name>.<statistic>, where db_name is the path of the database file, and they are removed once no listener uses the provider any more.
happy_eyeballs: The happy eyeballs sorting of a multi-address host’s address list now happens once when the address list is created or refreshed, instead of on every upstream connection attempt. The order in which connection attempts are made is unchanged.
http: The x-envoy-upstream-stream-duration-ms request header is now removed from external requests, in line with the other router timeout and retry headers, so it is only honored for requests from internal clients. Previously any client could use it to set the maximum upstream stream duration. This change can be reverted by setting the runtime guard envoy.reloadable_features.sanitize_upstream_stream_duration_header to false.
http: The HTTP connection manager now proactively drains the connections of a draining listener in the last third of the drain time, rather than only checking the drain state when a response is sent. This avoids idle connections and connections with upgraded (e.g. WebSocket) or CONNECT streams all being closed at the same time at the end of the drain sequence. Connections are not drained proactively when --drain-time-s is too short to leave room for it after the connection manager’s drain timeout. This behavioral change can be temporarily reverted by setting the runtime guard envoy.reloadable_features.use_connection_event_drain to false.
http: The HTTP filters of the connection manager are now handed the http.<stat_prefix>. scope of the connection manager as the stats prefix scope of their factory context, and the stats prefix that the filter factories read through ExtraFactoryContext::statsPrefixOr() is empty instead of that same http.<stat_prefix>. string. Stat names, tag extracted names and tags are unchanged: the prefix only moves from the name that each filter builds to the name of the scope the filter creates its stats in. The filters whose stats would not have survived that move unchanged keep creating them outside that scope, from the full stats prefix. This change can be reverted by setting the runtime guard envoy.reloadable_features.use_stats_prefix_scope_for_http_filter to false.
http: The QUERY request method, registered by RFC 10008, is now recognized by the HTTP/1 codec and is forwarded rather than rejected with a 400 (HPE_INVALID_METHOD). HTTP/2 and HTTP/3 have no method allowlist and already forwarded it. QUERY was also added to the method registry that restrict_http_methods enforces. Deployments that relied on Envoy rejecting QUERY at the edge will now see those requests routed. This behavioral change can be temporarily reverted by setting runtime guard envoy.reloadable_features.http1_allow_query_method to false. In addition, RFC 10008 Section 2 requires servers to fail a QUERY request whose Content-Type field is missing, so such a request is now rejected with a 400 and the response code detail query_missing_content_type. This applies to every downstream protocol, including HTTP/2 and HTTP/3 where these requests were previously forwarded. Consistency between the declared media type and the request content is left to the origin server.
http: The runtime guard envoy.reloadable_features.use_canonical_suffix_for_quic_brokenness now defaults to true. When enabled, the HTTP server properties cache uses configured canonical suffixes to share QUIC brokenness status across matching origins.
http: The values of the envoy.reloadable_features.match_headers_individually, envoy.reloadable_features.http2_include_cookies_in_limits and envoy.reloadable_features.http2_discard_host_header runtime features are now latched at header-matcher or codec-connection construction time instead of being looked up on hot code paths (per header field or per header match). Runtime overrides of these flags now take effect for newly created connections and newly loaded configurations rather than immediately for existing ones.
http2: Reserved flag bits on downstream HTTP/2 CONTINUATION frames are now ignored, per RFC 9113 §6.10, so they can no longer alias END_STREAM on the accumulated HEADERS frame and desynchronize Envoy’s stream state from the underlying codec. This change can be temporarily reverted by setting the runtime guard envoy.reloadable_features.http2_mask_continuation_flags to false.
jwt_authn: The jwt_authn HTTP filter now strips every configured forward_payload_header and claim_to_headers header name from the request before applying rules. Previously those headers were sanitized only inside the matched verifier, so paths that bypassed verification (empty requires, per-route disabled, or CORS preflight bypass) could forward client-supplied values upstream, and a request authenticated by one provider could retain spoofed payload/claim headers configured on another provider. Guarded by envoy.reloadable_features.jwt_authn_sanitize_payload_headers_filter_wide (default true).
mcp: Updated MCP message parsing to preserve dots in _meta field names, enabling reserved metadata keys such as io.modelcontextprotocol/protocolVersion used by MCP 2026-07-28 to be parsed correctly.
mcp: McpFilter no longer rejects any requests in PASS_THROUGH mode. Previously, the filter could reject requests in some cases when it failed to parse the request as an MCP request. Parse failures are now recorded in the filter’s dynamic metadata, under passthrough_reason.
mcp_transcoder: McpJsonRestBridgeFilter rejects configs that contain two tools with the same name, either both in the base config or both in the same route. This can be used to check for cases where the config provider does not fully validate incoming config.
oauth2: Route level OAuth2 configurations now register their SDS token and HMAC secrets with the init manager of the route configuration that owns them, so that route configuration is only published once those secrets are ready. Previously the secret subscriptions started immediately and the route configuration was published without waiting for them, so the filter could run against empty secrets until the first SDS update arrived. As a result, a route configuration referencing an SDS server that is slow or unreachable now takes correspondingly longer to warm up.
oauth2: The OAuth2 filter now respects user-configured retry_on in retry_policy. Previously, the value was overridden with 5xx,gateway-error,connect-failure,reset, so a configured retry_on had no effect on requests to the OAuth server. A retry_policy which does not set retry_on now retries nothing, rather than silently inheriting those four conditions. Controlled by runtime flag envoy.reloadable_features.oauth2_client_retries_respect_user_retry_on (defaults to true); set to false to preserve the old behavior.
quic: Promoted QUIC/HTTP3 from alpha to stable. This includes all QUIC extensions and upstream HTTP/3 support. HTTP/3 downstream was already considered production-ready, and HTTP/3 upstream is now also considered stable.
ratelimit: metadata: MetadataKey now supports accessing list elements via index. Path segments can now specify either a struct field name (key) or a list element index (index). This allows the ratelimit filter’s metadata descriptor action to access values extracted by grpc_field_extraction (which writes ListValue) without requiring a Lua bridge. For example, to access the first element of a list at envoy.filters.http.grpc_field_extraction.tenant_id, use path: [{key: tenant_id}, {index: 0}]. If the index is out of bounds or the value is not a ListValue, the lookup returns an empty Value (same behavior as accessing a non-existent key).
rbac: Fix: CVE-2026-73553

RBAC path matching (via PathMatcher and UriTemplateMatcher) now respects the route’s ignore_path_parameters_in_path_matching configuration. When enabled on a route, the RBAC filter will strip path parameters (everything after a semicolon in each path segment, e.g., transforming /admin;x=y/action;foo=bar to /admin/action) before evaluating the path match. This ensures path matching consistency between the Router and the RBAC filter, preventing authorization bypasses where an attacker could append path parameters to bypass RBAC rules while still being routed to the protected endpoint.

This behavioral change can be temporarily reverted by setting the runtime guard envoy.reloadable_features.rbac_respect_ignore_path_parameters to false.
rbac: Upstream HTTP filter stats are now correctly scoped under the parent’s stat prefix (http.<stat_prefix>.rbac.* for router filters, cluster.<name>.rbac.* for cluster filters). Guarded by runtime flag envoy.reloadable_features.upstream_http_filters_correct_stats_prefix (default true).
redis_proxy: The Redis proxy codec is stricter about malformed RESP wire input that was previously accepted silently: negative aggregate or bulk length headers other than the spec’s *-1 / $-1 null forms, integer lines carrying no digits, integers outside the signed 64-bit range, and messages exceeding new nesting-depth, cumulative-element, inline-command-element and scalar-token limits are now treated as protocol errors that close the connection. A single bulk string, blob error or verbatim string payload is additionally capped at 512 MiB, matching Redis’s default proto-max-bulk-len; deployments whose backends raise proto-max-bulk-len beyond that default are affected by this cap. Locally generated error replies also have ASCII control bytes replaced with spaces so attacker-influenced text cannot inject RESP framing. The remaining changes are only visible to peers sending non-conforming or abusive wire data.
reverse_tunnel: The reverse_tunnel network filter now fails closed when a configured identifier validation formatter (node_id_format, cluster_id_format, tenant_id_format) renders an empty string or the absent-value placeholder -. Previously such a render silently skipped the identity binding and accepted the claimed identifier. A present-but-empty x-envoy-reverse-tunnel-tenant-id header value is now rejected with a 400 error, matching the existing missing-header rejection and the empty-identifier guards already applied to node and cluster ids at socket registration.
reverse_tunnel: The downstream reverse tunnel socket interface (envoy.bootstrap.reverse_tunnel.downstream_socket_interface) now validates rc:// initiator listener addresses when they are resolved. The per-host connection count must be within [1, 1024], and each identifier (source node, cluster, and tenant ids and the remote cluster name) must be at most 255 bytes and a valid HTTP header value. An rc:// address with a count of 0 or above 1024, or an identifier that is too long or not a valid header value, now fails to load where it was previously accepted. Separately, a reverse connection to a remote cluster host that resolves to an EnvoyInternal address is now rejected before dialing, because the user-space I/O handle cannot be duplicated for the accepted tunnel.
server: Envoy builds using gperftools tcmalloc now honor memory_allocator_manager.bytes_to_release, releasing the configured number of bytes every memory_release_interval. Previously the setting was ignored. A memory_release_interval of zero, or shorter than one millisecond, now disables background release for both Google’s tcmalloc and gperftools tcmalloc. Builds using neither allocator now log a warning when bytes_to_release is non-zero.
server: Fixed container-aware CPU limit detection (#45410) not being enabled by default. The minimum of the cgroup CPU limit, CPU affinity, and hardware thread count, added in #40997 and documented as the default in the v1.37.0 release notes, was only applied when --cpuset-threads was set. It is now applied whenever --concurrency is not set, so worker threads are sized to the cgroup CPU limit in containerized deployments without requiring --cpuset-threads. Detection can still be disabled by setting ENVOY_CGROUP_CPU_DETECTION to false.
sockets: If Envoy sends an RST over a connection to/from an internal listener, the userspace socket’s IOHandle implementation will propagate the RST to the peer handle. This behaviour change can be disabled by setting the runtime guard envoy.reloadable_features.enable_send_rst_on_user_space_socket to false.
stats: Added the runtime guard envoy.reloadable_features.enable_stats_explicit_tags (default false). When set to true and the stats configuration carries no custom tags (empty stats_tags and use_all_default_tags left at its default of true), the stats store uses the tags supplied by the calling code (the explicit-tags logic) and propagates scope-level tags onto every stat, instead of re-parsing the flat stat name. The guard is evaluated once at startup. There is no visible change to users while the guard remains false.
stats: Stat-name construction no longer allocates when a join has at most one non-empty operand. Joining a name with an empty name produces bytes identical to that name, so TagStatNameJoiner and the HTTP response-code stat helpers now reference the non-empty name directly instead of allocating a byte-identical copy. The router, ext_authz and ratelimit all charge response-code stats with an empty prefix, so this removes four heap allocations per upstream response. The resulting stat names are unchanged.
tls: Added X25519MLKEM768 (hybrid post-quantum key exchange) to the default ECDH curves for both upstream and downstream TLS connections (including FIPS builds). This can be disabled by setting the runtime flag envoy.reloadable_features.pqc_default_ecdh_curves to false.
tracing: tracing: the OpenTelemetry tracer now populates the flags field on exported OTLP spans. The low 8 bits carry the W3C trace flags of the span (currently only the sampled bit), and bits 8 and 9 record whether the span’s parent context was remote, as defined by the OTLP specification. Previously the field was always 0, which OTLP consumers interpret as “trace flags not recorded”.
upstream: The runtime guard envoy.reloadable_features.coalesce_lb_rebuilds_on_batch_update now defaults to true. A thread-aware load balancer (for example RING_HASH or MAGLEV) rebuilds its factory state once at the end of a batch host update, from the single end-of-cycle member-update callback, instead of once per priority from the per-priority update callback. The rebuild still lands before the cluster manager posts the update to the worker threads, so this only removes the redundant per-priority rebuilds of a batch. This can be reverted by setting envoy.reloadable_features.coalesce_lb_rebuilds_on_batch_update to false.
wasm: The identity of a Wasm plugin (which plugin configurations share a single root context and thread-local plugin instance inside a Wasm VM) is now derived from the whole plugin configuration instead of from the plugin name and the traffic direction of the listener the plugin was configured on. Configurations that differ in any field other than vm_config no longer share an instance, and identical configurations now share one regardless of the traffic direction they are configured on.
watchdog: Configuring the envoy.watchdog.backtrace_action now causes Envoy to install a process-wide SIGUSR2 signal handler and to send SIGUSR2 to stuck threads in order to capture their backtraces. Deployments that rely on SIGUSR2 for other purposes should avoid enabling this action.

Bug fixes

Changes expected to improve the state of the world and are unlikely to have negative effects

access_log: Fixed a bug where omit_empty_values had no effect for json_format. Because the JSON formatter pre-serializes the template when loading the configuration, keys whose command operators evaluated to null were still emitted (for example {"key":null} instead of {}). When omit_empty_values is set, the JSON formatter now omits keys with null values, removes nested objects that become empty, and preserves empty arrays, matching the documented behavior. This behavioral change can be reverted by setting the runtime guard envoy.reloadable_features.json_formatter_omit_empty_values to false.
admin: Fix: CVE-2026-73546

Sanitize stat names before converting them to HTML. The change is guarded by runtime guard envoy.reloadable_features.sanitize_html_stats_names.
aggregate_cluster: Fixed a use-after-free crash in the aggregate cluster load balancer factory. The factory is shared with every worker thread but held a raw reference to its Cluster, which a CDS update can destroy on the main thread before a worker runs its queued cluster-add callback and dereferences it, typically at startup. The factory now copies the state it needs to create the load balancer instead of referencing the Cluster.
api: Fixed incorrect realm value in BasicAuth filter. Added a realm field to the basic_auth HTTP filter, allowing browsers to correctly cache credentials across paths. When unset, the previous behavior is preserved, the realm is derived from the request URI.
api_key_auth: Fixed a crash in the api_key_auth HTTP filter when hide_credentials is enabled with a query parameter key source and a request without a :path header (for example a CONNECT request) is authenticated via another key source. The filter now skips query-string rewriting when there is no :path.
aws: Fixed a data race in the AWS credentials file provider (CredentialsFileCredentialsProvider) that could corrupt the heap and crash Envoy when watched_directory was configured. With a watched directory the cached credentials were refreshed on every getCredentials() call, and concurrent worker threads wrote the cached Credentials and last_updated_ members without synchronization. The cached state is now guarded by a mutex.
aws: Fixed a race where the route level configuration of the AWS request signing and AWS Lambda filters could be destroyed on a worker thread when an RDS update replaced the route configuration.
aws: Fixed an initialization hang when an extension that gates server initialization sends an HTTP callout to a cluster whose AWS request signing filter resolves credentials asynchronously.
cache: Fixed an assertion failure in the cache_v2 filter caused by sendHeaders() calling firstBytePos() for suffix range requests such as Range: bytes=-2.
cache: http cache: fixed cache and cache_v2 to ignore case in Cache-Control directive names, as required by RFC 9111 section 5.2. Responses with Private or No-Store are now rejected by the cache, just like private or no-store.
cares: Changes the default value of envoy.restart_features.shared_cares_dns_resolver to false. This disabled the shared dns resolver that can cause a race when createDnsResolver() is called from a workerthread in the DnsFilter. Do not turn this back on until this bug is fixed if DnsFilter is used.
composite: Fixed a race where the route level configuration of the composite filter could be destroyed on a worker thread when an RDS update replaced the route configuration. The match tree, and the delegated filter configurations its actions own, are now released on the main thread.
composite_cluster: Fixed composite clusters failing requests with 503 no_healthy_upstream when the sub-cluster selected for an attempt has no host available, for example because its endpoint list is empty or because all of its hosts have been ejected. The composite cluster now fails over to the following sub-clusters in the configured list within the same attempt. This behavior can be reverted by setting the runtime guard envoy.reloadable_features.composite_cluster_skip_clusters_without_hosts to false.
config: Fixed a bug in on-demand xDS where a resource requested after the initial subscription could be silently dropped instead of delivered. An on-demand update only updated the subscription sent to the management server, not the internal watch routing, so the server’s response for the requested resource matched no watch and was discarded. This affected on-demand cluster discovery (ODCDS) when requesting a second, different cluster, and VHDS virtual hosts under a route configuration. On-demand requests now also register watch routing, so these responses are delivered.
cpu_utilization: Fixed the envoy.resource_monitors.cpu_utilization monitor in CONTAINER mode when Envoy shares the host cgroup namespace, for example in a privileged container. /sys/fs/cgroup is then the cgroup v2 root, which has no cpu.max, so detection failed and config initialization aborted the server. The monitor now resolves its own cgroup from /proc/self/cgroup and /proc/self/mountinfo and reports container rather than host usage. cgroup v2 detection is keyed on cpu.stat: an absent cpu.max means no CPU limit and an absent cpuset.cpus.effective falls back to the CPU affinity count, while a file that exists but cannot be read fails the sample.
credential_injector: Fixed a bug in the credential_injector OAuth2 client_credentials provider where the token request body was percent-encoded with a generic URI-component character set instead of the application/x-www-form-urlencoded algorithm. A literal + in client_id, client_secret, scope, or a custom token_endpoint parameter was left unescaped and silently decoded by the token endpoint as a space, corrupting the credential. This most commonly affected base64-encoded IdP-issued client secrets, which contain + roughly half the time.
credential_injector: Fixed a bug where a credential loaded from a file-based generic secret was injected into the request header verbatim, including any trailing newline commonly present in secret files. Since HTTP header values cannot contain CR/LF, this produced an invalid header and the request failed. Trailing CR/LF characters are now stripped from the credential before injection, and a credential consisting only of CR/LF characters is treated as missing.
dns_resolver: Fixed a use-after-free in the Hickory DNS resolver. A resolution completing after the resolver was destroyed dereferenced the freed resolver before checking whether it was still alive, for example when a module task outlived the shutdown timeout. The resolver now hands the module a never-reused token instead of its own pointer and resolves that token under a lock before touching any resolver state, so a late completion is dropped safely.
dns_resolver: Fixed the c-ares resolver reporting a successful empty result when one address family returned no records and the other lookup failed in AUTO or V4_PREFERRED mode. Such failures no longer remove existing hosts from strict DNS clusters. This change can be reverted by setting the runtime guard envoy.reloadable_features.cares_dual_resolution_preserve_failure to false.
dns_resolver: Fixed the c-ares resolver to preserve a reentrant query when it reuses the completing query’s DNS transaction ID. Previously, the old query could remove the new query’s ID mapping and permanently stall DNS refresh for the affected cluster.
dns_resolver: Updated the Hickory DNS resolver crates from 0.26.1 to 0.26.3, which fixes DNSSEC validation, denial of service, and DNS message parsing vulnerabilities.
dynamic_forward_proxy: Fixed a use-after-free crash in the DNS cache manager: the server-wide DnsCacheManager singleton no longer retains the stats scope of the listener or filter chain that first created a DNS cache, which could be freed before a later cache miss for a new cache name dereferenced it.
dynamic_modules: Fixed a bug where the envoy_dynamic_module_on_http_filter_per_route_config_destroy hook of a dynamic module could run on a worker thread, and the module could be unloaded there, when an RDS update replaced the route configuration.
dynamic_modules: Fixed a crash in the dynamic modules Go SDK where a panic raised inside a module hook crossed the cgo export boundary and aborted the whole process. Every Go SDK export now recovers a panic at the ABI boundary, logs it at error level, and returns a fail-closed value, mirroring the Rust and C++ SDK panic barriers.
dynamic_modules: Fixed a crash where a panic inside a Rust dynamic module bootstrap stats visitor passed to iterate_counters or iterate_gauges unwound across the C ABI boundary and aborted Envoy. The SDK now catches the panic, stops the iteration, and logs the failure so the process stays up.
dynamic_modules: Fixed a use-after-free crash in the dynamic modules HTTP filter. An event hook that ends the stream, for example envoy_dynamic_module_callback_http_filter_recreate_stream, tears the filter chain down on the module’s own stack, which freed the in-module filter the hook was still running on. The in-module filter is now destroyed from the dispatcher’s deferred deletion list, so envoy_dynamic_module_on_http_filter_destroy runs once every other event hook has returned. The callbacks that need the torn-down stream no longer dereference it, and HTTP callouts started after the teardown are refused instead of outliving the filter.
dynamic_modules: Fixed a use-after-free in the dynamic modules network filter. A module that read the network read buffer outside on_read, for example from on_scheduled or on_http_callout_done, could read freed stack memory when the buffer delivered to on_read was a transient injected buffer from inject_read_data, the network ext_proc filter, or the tcp_bandwidth_limit filter. Deferred access now resolves the connection read buffer, which stays valid for the connection lifetime.
dynamic_modules: Fixed crashes and silent no-ops when a dynamic module returns a null object from an on_*_new hook. A null UDP listener filter config is now rejected at config load, a null cluster load balancer no longer registers the host membership callback, and a null bootstrap extension is rejected at load. The cluster constructor also checks the base construction status before running the module hook.
dynamic_modules: dynamic modules: added an exception barrier to the C++ SDK so an exception thrown from a module hook is caught at the ABI boundary and fails closed instead of aborting the process. This mirrors the Rust SDK panic barrier.
dynamic_modules: dynamic modules: fixed a bug in the C++ SDK where a filter stopped receiving onResponseHeaders, onResponseBody and onResponseTrailers after any local reply on the stream, including local replies the module did not send. A module could not observe or modify the response for a direct_response route, a local reply from another filter, or an Envoy-generated error. The Rust and Go SDKs were unaffected.
dynamic_modules: dynamic modules: fixed a bug where removing cluster hosts that were added at a priority above zero did not republish that priority, so worker load balancers kept routing to the removed endpoints. Hosts added through the cluster host APIs now also record their real priority, which keeps the cross priority host map consistent when they are removed.
dynamic_modules: dynamic modules: fixed a bug where the Rust SDK cluster add_hosts methods passed weights, localities, and metadata to the host without checking their lengths against the address count. A module could trigger an out of bounds read in the host by passing mismatched or ragged slices. The SDK now returns None before crossing the ABI when the slice lengths are inconsistent.
dynamic_modules: dynamic modules: fixed a bug where the built-in Hickory DNS resolver aborted the process when it could not create its Tokio runtime or resolver, for example under resource exhaustion. The module now reports the failure and Envoy rejects the configuration with an error instead of aborting.
dynamic_modules: dynamic modules: fixed a bug where the network and listener filters could reenter a module when an HTTP callout completed inline, before the async client returned a request handle. The callout success and failure callbacks now skip an inline completion, which the module already observes through the callout return code.
dynamic_modules: dynamic modules: fixed a compatibility regression where a tracer module built against an SDK that predates the reserve_tags hook failed to load, because the host resolved that hook as mandatory. The reserve_tags hook is now resolved optionally, and the host skips the call when the module does not define it.
dynamic_modules: dynamic modules: fixed a main thread deadlock in the bootstrap stats iterate callbacks. They invoked the module iterator while holding the stats allocator lock, so a stats call from the iterator re-entered the non-recursive lock and hung the process. The counters and gauges are now snapshotted before the iterator runs, so it runs without the lock held, and a Stop return now ends iteration.
dynamic_modules: dynamic modules: fixed a startup crash where a bootstrap extension configured with a remote module source and nack_on_cache_miss dereferenced the cluster manager before it was created. Bootstrap extensions now reject a remote module source at config load, since the cluster manager needed to fetch it does not exist yet at bootstrap time.
dynamic_modules: dynamic modules: fixed a use-after-free in the HTTP filter. Consecutive typed filter-state or host metadata getter calls in the same event hook shared a single scratch buffer, so an earlier returned view was freed by a later call. Each getter now appends to a per-hook scratch that stays valid until the event hook returns.
dynamic_modules: dynamic modules: fixed a use-after-free in the listener filter. Consecutive typed filter-state getter calls in the same event hook shared a single scratch buffer, so an earlier returned view was freed by a later call. Each getter now appends to a per-hook scratch that stays valid until the event hook returns.
dynamic_modules: dynamic modules: fixed a use-after-free in the network filter. Consecutive typed filter-state getter calls in the same event hook shared a single scratch buffer, so an earlier returned view was freed by a later call. Each getter now appends to a per-hook scratch that stays valid until the event hook returns.
dynamic_modules: dynamic modules: fixed a use-after-free in the stats sink where the histogram name passed to on_histogram_complete was serialized into a single thread local buffer. A module that recorded a histogram from inside on_histogram_complete re-entered the sink on the same thread and overwrote the name the outer callback was still reading. The outermost call now keeps the shared buffer while a re-entrant call uses its own buffer.
dynamic_modules: dynamic modules: fixed a use-after-free in the upstream HTTP to TCP bridge. A module that sent a complete response from an event hook before the downstream request finished caused the router to reset and destroy the bridge while the module hook was still running. The bridge now applies a terminal response after the hook returns.
dynamic_modules: dynamic modules: fixed a use-after-free where the HTTP and network filter socket option byte value getters returned a view into a std::vector that relocated its elements when a later option was stored, dangling views the ABI promises stay valid for the filter lifetime. The options are now kept in a std::deque with stable element addresses, and the getters return the latest value set for an option rather than the first.
dynamic_modules: dynamic modules: fixed a use-after-free where the cluster load balancer typed filter state getter returned a view backed by a single thread local string. A module that read two keys during a host selection callback saw the first view clobbered by the second read. Serialized values are now retained for the whole host selection callback and cleared when it returns.
dynamic_modules: dynamic modules: fixed missing scalar accessors for xDS cluster names, listener directions, and filter chain names, and preserved empty route names as available values.
dynamic_modules: dynamic modules: fixed request and response size and gRPC status scalar attributes in HTTP-aware contexts.
dynamic_modules: dynamic modules: fixed the socket option callbacks so a level, name, or value that does not fit in an int is rejected instead of truncated and applied as a different option. This affects the HTTP, network, and listener filter socket option callbacks.
dynamic_modules: dynamic modules: hardened the handling of two module-provided values that Envoy reads after the producing hook returns. The on_program_init version string is now copied out of module memory right away, and the ABI documentation for the version string and the cert validator digest buffer now states that they must stay valid until Envoy reads them immediately after the hook returns.
ext_authz: Fix: CVE-2026-50572

Fixed UAF when ext_authz over HTTP causes request to be rejected.
ext_authz: Fix: CVE-2026-73547

Fixed abnormal process termination when Envoy calls ext_authz service with requests without URI path (i.e. CONNECT).
ext_authz: Hardened the ext_authz filter to avoid a crash when building the metadata context for a stream that has no downstream connection (for example when the filter runs in an upstream filter chain driven by the async client). Connection metadata is now included only when a connection is present.
ext_proc: Fixed multiple lifetime bugs in the external processing (ext_proc) filter and the underlying gRPC async client that could lead to use-after-free or double delivery of callbacks. The gRPC async client now holds an optional reference to its stream callbacks and drops it once the stream is cleaned up or the owner detaches via waitForRemoteCloseAndDelete(), so a stream that outlives its callbacks (for example while awaiting remote close) no longer invokes callbacks on freed memory. Re-entrant resets during stream initialization are guarded so remote close is not notified (and the tracing span not finished) twice when the cluster is missing or stream creation fails synchronously, and half-close/cleanup no longer dereference a stream that was never established. The ext_proc ThreadLocalStreamManager and ProcessorStreamImpl now close any still-open streams on destruction to avoid dangling references into the underlying gRPC stream.
ext_proc: Hardened the ext_proc filter to avoid a crash when logging stream info for a stream whose upstream info is null. The upstream host is now recorded only when upstream info is present.
filter_chain: Fixed a race where the route level configuration of the filter chain filter could be destroyed on a worker thread when an RDS update replaced the route configuration. The embedded filter chain, and the filter configuration providers it owns, are now released on the main thread.
geoip: Fixed a bug in the MaxMind geolocation provider where a configuration that had previously been used and then torn down could return a null driver instead of building a new provider.
geoip: Fixed a bug in the MaxMind geolocation provider where a database path that could not be opened, for example a path that does not exist, crashed Envoy during configuration load. Such a configuration is now rejected instead.
geoip: Fixed a bug in the geoip filters where lookups would be attempted on non-IP addresses, such as internal or Unix domain socket addresses. The filters now skip the lookup when no IP address is available and increment the skipped counter.
geoip: Fixed a bug where the MaxMind geolocation provider ignored the stat prefix of every listener but the first.
grpc_http1_reverse_bridge: Fixed a crash (SEGFAULT) in the grpc_http1_reverse_bridge filter when withhold_grpc_frames is enabled without response_size_header and the upstream response body exceeds the downstream HTTP/2 stream flow control window. The filter now uses the upstream Content-Length header to stream the response incrementally instead of buffering and releasing it all at once.
http: Enforced a 10-second minimum idle duration before terminating idle HTTP/3 sessions when the envoy.overload_actions.close_idle_http_connections overload action is saturated.
http: Fix: CVE-2026-73548.

Fixed a vulnerability where payload sent before a generic HTTP upgrade was accepted could be interpreted as a pipelined HTTP/1 request and poison a shared upstream connection. Generic upgrade payload is now paused until the upstream accepts the upgrade. This change can be temporarily reverted by setting envoy.reloadable_features.http_pause_generic_upgrade_request_body to false.
http: Fixed a bug in the HTTP cache v2 filter where body and trailer reset callbacks could run on the wrong worker thread when a cache operation failed or a body request was out of range.
http: Fixed a bug where malformed CONNECT request lines without an authority could cause the legacy HTTP/1 parser to encode a 400 Bad Request response using HTTP/1.0. Envoy now rejects these requests without downgrading the response protocol.
http: Fixed a bug where response metadata added by HTTP encoder filters could be dropped when a later encoder filter sent a direct local reply before final response headers were encoded to the codec. Saved response metadata is now flushed before the local reply ends the stream. This behavior can be temporarily reverted by setting the runtime guard envoy.reloadable_features.direct_local_reply_flush_saved_response_metadata to false.
http: Fixed a request/response body data-loss bug in the HTTP filter manager. When a filter stopped iteration on headers (for example a wasm filter with allow_on_headers_stop_iteration, which maps to a single-iteration stop rather than StopAllIterationAndWatermark), resumed asynchronously, and then on a subsequent body frame moved that frame into the filter-manager buffer via addDecodedData()/addEncodedData() before returning Continue, the now-empty frame was forwarded down the chain and the buffered bytes were silently dropped. This corrupted large streamed request bodies (for example one 16 KiB chunk lost when chained with an ext_proc filter in FULL_DUPLEX_STREAMED mode). The just-buffered data is now forwarded instead of the empty frame. This behavioral change can be reverted by setting the runtime guard envoy.reloadable_features.filter_manager_forward_added_data_on_continue to false.
http: Fixed the HTTP cache filters so that a response s-maxage directive implies proxy-revalidate (RFC 9111) and is not served stale when the request allows max-stale.
http: Fixed the cache_v2 filter incorrectly rewriting non-200 responses (for example 404) to 206 or 416 when the request included a Range header. Range is now applied only to 200 responses. See RFC 9110 Section 14.2.
http: Fixed the cache v2 filter inserting responses into the cache when the request contains Cache-Control: no-store. The ignore_request_cache_control_header option continues to allow these responses to be cached when enabled.
http: Fixed the custom response filter so %LOCAL_REPLY_BODY% in a local response policy’s body_format receives the existing local reply body when the policy does not configure its own body.
http: Fixed unbounded memory growth caused by TlsCachingDateProviderImpl::onRefreshDate() calling SlotImpl::set() every 500ms on the main thread while workers had not started, leaving Date update callbacks queued in their dispatchers. This occurred when xDS was unavailable during startup with initial_fetch_timeout: 0s. Each thread now initializes its Date cache immediately and refreshes it with a local timer. See issue #31561.
http2: Fix: CVE-2026-73513

Fixed abnormal process termination when Envoy receives trailers without the END_STREAM flag over HTTP/2 protocol.
http2: Fixed HTTP/2 connection write-buffer low-watermark callback delivery when a callback reentrantly encodes data. Previously, reordering the active-stream list during callback delivery could notify one stream twice and skip another, leading to a stalled response.
http2: Fixed abnormal process termination when an HTTP/2 peer sends HEADERS or DATA frames on a stream that it has already ended. Such frames are now rejected with a codec protocol error, as required by RFC 9113 Section 5.1, instead of being dispatched to a decoder that may already have been destroyed. This behavior can be reverted by setting the runtime guard envoy.reloadable_features.http2_reject_frames_after_end_stream to false.
http2: Fixed an integer overflow in HTTP/2 codec stream flow control accounting where unconsumed_bytes_ wrapped around when reads were disabled on a stream receiving over 4 GiB of data.
http2: Fixes CVE-2026-73550

Account for the length of dropped Host headers in HTTP/2 request header map size and count limits. Host headers are dropped when they match HTTP/2 `:authority header.

This behavioral change can be reverted by setting the runtime guard envoy.reloadable_features.http2_track_size_of_dropped_host_header to false.
http3: Fix: CVE-2026-48521

Fixed abnormal process termination when Envoy is configured to automatically select a protocol with upstream server based on ALPN and the server uses HTTP/3.
http3: Fix: CVE-2026-73512

Fixed UAF when Envoy receives specifically timed sequence of HTTP/3 frames.
http3: Fixed a bug where HTTP/3 load shed points envoy.load_shed_points.http3_server_go_away_on_dispatch and envoy.load_shed_points.http3_server_go_away_and_close_on_dispatch were evaluated on every UDP packet rather than only on new stream creation. This behavioral change can be reverted by setting the runtime guard envoy.reloadable_features.http3_fix_goaway_loadshed_point to false.
ip_tagging: Fixed a bug in the ip_tagging filter where stats could be published under another listener’s stats prefix. Filter configs that load tags from the same ip_tags_datasource file share same cached provider and could end up publishing stats under the wrong listener’s stats prefix.
jwt_authn: Fixed a bug where a claim_to_headers entry whose claim could not be resolved in the JWT payload was dropped without any trace. Such an entry is now logged at debug level.
jwt_authn: Fixed the remote JWKS fetch span so it honors the parent span’s sampling decision. Previously the JWT Remote PubKey Fetch span was always sampled because the async client request options left the sampling decision at its default; the decision is now left unset so the span inherits the parent span’s sampling decision.
listener: Fixed a crash at startup when a UDP or QUIC listener was configured with bind_to_port: false. This combination was never functional and is now rejected at configuration load with a validation error.
listener: Fixed a use-after-free when a listener filter calls continueFilterChain from inside its own onAccept, onData, or onClose hook. The re-entrant continuation cleared the accept filter list under the running filter, leaving a dangling iterator and a null socket. The continuation is now deferred and applied after the hook returns. This behavior is guarded by envoy.reloadable_features.listener_filter_reentrant_continue_guard.
local_ratelimit: Fixed a race where the share provider manager owned by a local rate limit route level configuration could be destroyed on a worker thread. The manager is an unpinned singleton that owns a cluster membership callback handle, so a route level configuration may own the last reference to it and release it when an RDS update replaces the route configuration. It is now handed to the main dispatcher alongside the rate limiter, which was already posted there.
lua: Hardened the lua filter to avoid a crash in connectionStreamInfo() when the downstream connection is absent. The call now returns nil to the script instead.
mcp: Fixed MCP JSON parser path tracking for object-valued keys containing ., which could cause subsequent fields such as params.name, params._meta, or id to be omitted from extracted metadata.
mcp: Fixed a memory usage issue in the MCP JSON-RPC parser by optimizing node allocation for unneeded fields.
mcp_json_rest_bridge: Fixed a path-traversal issue in the mcp_json_rest_bridge HTTP filter where a path-template variable’s value (taken from attacker-controlled tool-call arguments) was installed verbatim into the upstream request :path, so a value such as ../../admin/secrets produced raw path traversal. Traversal segments (. / ..) are now rejected for every template variable, and a “simple” variable (for example {id}) additionally has / percent-encoded to confine it to a single path segment. Variables with an explicit pattern such as {name=projects/*} may still legitimately span multiple segments.
mcp_json_rest_bridge: mcp_json_rest_bridge: Fixed a bug where headers-only upstream responses (e.g., HTTP 204 No Content) were passed through to MCP clients without a JSON-RPC response body, causing MCP SDK timeouts or exceptions. The filter now synthesizes a valid JSON-RPC response: an empty ToolResult for tools/call requests and a server error for tools/list requests.
oauth2: Fixed a bug in the OAuth2 filter where, when refreshing an access token with BASIC_AUTH, the refresh token was percent-encoded with the default reserved character set instead of the application/x-www-form-urlencoded one used on every other token endpoint request. A refresh token containing +, /, =, &, ? or : was therefore sent to the token endpoint incorrectly encoded, and the refresh failed.
oauth2: Fixed a crash in the oauth2 HTTP filter when a request without a :path header (for example a plain CONNECT tunnel request) reached the filter. The filter now rejects a request that has no :path with a 400 (Bad Request) local reply.
on_demand: Fixed the on_demand filter waiting forever, or requesting the same virtual host in a loop, when on-demand VHDS delivers the requested virtual host but none of its routes match the request. The request now continues to the router, which replies with a 404. This behavior can be temporarily reverted by setting the runtime guard envoy.reloadable_features.on_demand_vhds_require_route_match to false.
opentelemetry: Fixed the OpenTelemetry access loggers (both the gRPC and HTTP variants) ignoring configured formatters when building custom_tags. Previously a custom tag whose value used a formatter extension command failed with Not supported field in StreamInfo, even though the same command worked in body and attributes. The configured command parsers are now passed through to custom-tag creation.
orca: Fixed parsing of ORCA load reports sent as base64-encoded binary headers without padding, such as the endpoint-load-metrics-bin header emitted by grpc-go. These reports were previously rejected with unable to decode ORCA binary header value. This change can be reverted by setting runtime guard envoy.reloadable_features.orca_accept_unpadded_base64 to false.
postgres_proxy: Fixed the postgres_proxy filter forwarding incomplete initial message bytes to upstream when the message arrives in multiple TCP segments, causing PostgreSQL to reject the connection.
proxy_protocol: Fixed the upstream proxy_protocol transport socket ignoring added_tlvs entries that use format_string instead of a static value. These entries are now evaluated against the upstream connection stream info and emitted in the proxy protocol v2 header. An entry that sets both value and format_string, or neither, is now rejected at configuration load. This behavior can be reverted by setting the runtime guard envoy.reloadable_features.proxy_protocol_added_tlvs_format_string to false.
quic: Fix: CVE-2026-73549

Fixed a crash when handling scoped IPv6 addresses in QUIC client connection and Original Dst cluster.
quic: Fixed QUIC connection info to serve peer certificate details through the same code path as TLS connections. The peer certificate issuer digest and issuer serial number (used for example by the Lua filter) and the raw PEM-encoded peer certificate (used for example by CEL attributes) are now populated on QUIC connections, and peer certificate details remain available after the SSL object has been released when envoy.reloadable_features.quic_enable_reset_ssl_after_handshake is enabled.
rds: Fixed a bug where an RDS update carrying an invalid VHDS configuration was applied only halfway. The new route configuration was recorded before the VHDS subscription it configures was created, so when creating that subscription failed the update was rejected after the recorded state had already moved on: the admin /config_dump endpoint reported the rejected route configuration while the workers kept serving the previous one.
redis_proxy: Fixed a use-after-free in the Redis cluster CLUSTER SLOTS discovery. A cluster refresh (periodic resolve timer or DNS update) that arrived after CLUSTER SLOTS completed but while the zone-discovery INFO requests it triggered were still in flight could start a second discovery, overwrite the in-flight callbacks and free memory still referenced by the outstanding requests.
redis_proxy: Fixed an out-of-bounds read in the Redis inline-command decoder when parsing a quoted string whose hex escape (\xNN) appears at the very start of the token.
redis_proxy: Fixed an out-of-bounds read in the redis_proxy network filter when a transaction WATCH command (or a transaction’s first command) is received without a key argument. Such a request is now rejected with a wrong-number-of-arguments error.
redis_proxy: Fixed use-after-free crashes and resource leaks when a Redis cluster is removed. The discovery session could outlive the cluster (its discovery clients hold a reference to it) with the resolve timer still armed, and in-flight hostname resolutions and zone-discovery INFO requests were never cancelled. Cluster teardown now explicitly shuts the discovery session down, cancelling all in-flight discovery work and closing the discovery connections.
reverse_tunnel: Fixed a bug in the reverse tunnel downstream socket interface (envoy.bootstrap.reverse_tunnel.downstream_socket_interface) where handshake additional_headers values that use a ThreadLocal-backed substitution formatter (such as %FILE_CONTENT%, or secret/SDS-backed formatters) resolved to an empty string on the worker thread that assembles the handshake request. The handshake formatters were built in the bootstrap extension constructor, which runs before the worker threads register with the ThreadLocal system, so the formatter providers’ thread-local slots were never populated on the workers. The formatters are now built in onServerInitialized(), after the workers are registered, so their values propagate to every worker thread.
reverse_tunnel: Fixed a bug in the reverse tunnel downstream socket interface (envoy.bootstrap.reverse_tunnel.downstream_socket_interface) where handshake additional_headers values that use a substitution formatter (such as %FILE_CONTENT%, or secret/SDS-backed formatters) were sent as the raw, unsubstituted template on every reverse connection. The reverse connection listen socket snapshots the handshake formatters when it is created, which can happen before onServerInitialized() builds them; that null snapshot is then reused for every re-dial, so the handshake fell back to emitting the literal additional_headers value. The handshake headers are now resolved from the live bootstrap extension when the request is assembled, so post-initialization dials substitute the value correctly.
reverse_tunnel: Fixed a race where removing or draining a reverse-connection listener could trigger a new outbound handshake. Listener teardown now destroys the retry timer on its worker before closing the socket.
reverse_tunnel: Fixed a segfault in the reverse-tunnel initiator HTTP/1 handshake when the upstream rejected the handshake with a bodied response (for example 403 or 429). decodeHeaders() previously closed the connection while Http1::ConnectionImpl::dispatch() was still parsing the response body. The wrapper is now deferred-deleted and shutdown() closes the connection after dispatch returns.
reverse_tunnel: Fixed a use-after-free where removing a reverse-tunnel remote host (for example after a cluster host-map update) immediately destroyed in-flight handshake RCConnectionWrapper objects. Wrappers are now shut down (clearing the handshake read-filter back-pointer and HTTP/1 codec) and deferred-deleted on the worker dispatcher, matching the normal connection-done path, so a late handshake read cannot call into Http1::ConnectionImpl::dispatch() on a freed object.
reverse_tunnel: Fixed a use-after-free where reverse-tunnel listener stop left in-flight handshake RCConnectionWrapper objects alive until main-thread destruction. Worker resetFileEvents() now shuts down those wrappers and deferred-deletes them.
reverse_tunnel: Fixed reverse tunnel initiator listener handling in the downstream reverse tunnel socket interface. The listen handle now returns a fresh, unstarted handle from duplicate() rather than a raw file descriptor dup, so every worker under reuse_port dials and an LDS update of the reverse connection listener no longer removes the listener for the drain window. Tunnels still queued for accept() are closed on the worker that owns them when the listener stops, and a tunnel handle disposed without an explicit close, for example after a listener filter timeout, now releases its tunnel key so the host is redialed.
reverse_tunnel: Fixed several reverse tunnel handshake defects in the reverse_tunnel network filter. The HTTP/1 codec stats are now owned by the filter configuration rather than a per-dispatch local, removing a use after free that a peer could reach by pipelining handshakes. The handshake connection is now detached only after the acceptance response reaches the wire, so the duplicated tunnel socket is its single reader. A handshake that cannot be registered is answered with 503 instead of 200 over a dropped socket. A handshake request that carries a body is rejected with 400, and a handshake that does not complete within handshake_timeout (default 10s) is closed. The filter now requires the upstream reverse tunnel socket interface bootstrap extension at configuration load.
reverse_tunnel: Hardened reverse tunnel initiation in the downstream reverse tunnel socket interface. A dial that opens the connection but never receives a handshake response is now closed after the new handshake_timeout (default 15s) rather than holding the reverse tunnel slot forever. A handshake that fails synchronously or whose connection closes before completion is now terminal and installs backoff, and backoff is cleared only on a verified success. The requested host is selected strictly, so an unhealthy host is no longer dialed under another host’s key. Bytes the responder coalesces with the handshake response are carried into the accepted tunnel and replayed before the socket, and a queued connection that closes before accept() releases its tunnel key. Terminal handshake states are no longer stored per host, bounding memory under connection churn.
router: Fixed a bug where response_headers_to_remove was not applied to proxied upstream 1xx informational responses (such as 100 Continue and 103 Early Hints) when proxy_100_continue was enabled. This behavioral change can be temporarily reverted by setting the runtime guard envoy.reloadable_features.response_headers_to_remove_on_1xx to false.
router: Fixed a bug where shadowed (mirrored) requests did not honor dynamically-set subset load balancer metadata match criteria. Previously only the static route-level metadata_match was forwarded to the shadow cluster, so subset selectors set at runtime (for example via the header-to-metadata filter writing envoy.lb dynamic metadata, or connection-level envoy.lb metadata) were ignored and the shadow request could be routed to hosts outside the intended subset. The shadow stream now inherits the downstream request’s envoy.lb dynamic metadata (request-level merged over connection-level), matching the main request’s subset selection. This behavior can be temporarily reverted by setting the runtime guard envoy.reloadable_features.shadow_policy_inherit_dynamic_metadata to false.
router: Fixed a use-after-free when upstream_http_filters are configured via config_discovery (ECDS). The router held the upstream filter config provider manager, an unpinned singleton, only for the duration of its constructor, while the ECDS subscriptions it created retain a raw reference to that manager and dereference it when they are destroyed. If no cluster was keeping the singleton alive at the time the router configuration was built – for example a statically configured listener with CDS-supplied clusters – the manager was freed immediately and the subscriptions were left holding a dangling reference for the lifetime of the process. The router now retains the manager for as long as it owns the providers, matching how ClusterInfoImpl and the UDP proxy already hold it. The composite filter was hardened in the same way for dynamic_config actions.
safe_regex: Fix CVE-2026-73552

Switch safe_regex charset mode from UTF-8 to Latin1. HTTP headers are not UTF-8 encoded and must use Latin1 charset for regex expressions. This behavioral change can be temporarily reverted by setting runtime guard envoy.reloadable_features.re2_use_latin1_mode to false.
sds: Fixed a bug where on-demand SDS would start xDS subscriptions repeatedly triggering the initial fetch timeout. Fixed a bug where warming and non-warming (prefetch) SDS could incorrectly trigger each others’ readiness.
sds: Fixed a race where a generic secret provider shared with worker threads could be destroyed on a worker thread.
server: Fixed cgroup v1 CPU controller detection to match cpu as an exact controller name instead of matching controller names such as cpuset or cpuacct.
server: Fixed container-aware CPU limit detection for cgroup mounts whose mountinfo root is not /.
sockets: Fixed a crash in the TCP, HTTP, gRPC, and Thrift active health checkers that could occur when the upstream health check connection could not be created, for example when the configured Linux network namespace (network_namespace_filepath) became unavailable at runtime. The health check now reports a network failure instead of dereferencing a null connection.
sockets: Fixed hot restart dropping listeners bound in a network namespace whose network_namespace_filepath can no longer be opened. The new process now asks the parent for the existing listen socket before entering the network namespace, so a listener that is still serving in the parent is inherited even when the namespace path was removed after the socket was bound. Creating a new listener in a missing network namespace still fails.
sockets: Fixed the network namespace switch used for listeners and client connections with a network_namespace_filepath to return the calling thread to its own original namespace. It previously restored the main thread’s current namespace, so a worker thread creating a connection while the main thread was inside another namespace (for example creating a listener or a health check connection there) could be left in that namespace until its next switch, creating its sockets in the wrong namespace in the meantime.
stats: Fixed a bug introduced in 1.38.0 that exported Prometheus native histogram bucket indices one bucket too low, causing consumers to decode observations into lower value ranges and underreport quantiles.
stats: Fixed a bug where stats created with programmatic tags (via *FromStatNameWithTags) lost their tags across a hot restart. Such stats inline their tag values into the stat name, but the hot restart stat merge re-created them from the name alone with no tags, so after a restart the tag values collapsed into the metric name and the labels became empty (for example as exported to Prometheus). The parent now transmits the tag-extracted name and tag values for these stats, and the child re-creates them with identical labels. This behavior can be temporarily reverted by setting the runtime flag envoy.reloadable_features.hot_restart_propagate_stat_tags to false.
tcp_proxy: Fixed a stream leak in TCP proxy configured with tunneling_config routed through an internal listener. A full peer close on a user-space socket surfaced as a plain end-of-stream, which a half-close-enabled connection treats as a read half-close, so the upstream tunnel (e.g. HTTP CONNECT) stayed pinned until the upstream idle timeout. A full peer close is now reported the way the kernel reports a reset: reads drain pending data first and then return ECONNRESET. A peer shutdown(WR) half-close is unchanged and real OS sockets are unaffected. Guarded by the existing runtime feature envoy.reloadable_features.enable_send_rst_on_user_space_socket, which already gates turning a peer-side disconnect into a reset on the read side.
thrift_proxy: Fixed a 32-bit integer overflow in the thrift_proxy lax (non-strict) binary protocol decoder. A message name length of 0xFFFFFFF7 or greater wrapped the insufficient-data check in readMessageBegin and raised a spurious decode error that closed the downstream connection. The check is now performed in 64-bit arithmetic and the decoder waits for more data instead, matching the strict binary protocol.
tls: Fixed TLS session ID generation so that every CA certificate in the trusted CA bundle, not just the first one, contributes to the digest that keys resumable sessions. Previously a change to any CA after the first – for example rotating or removing a trust anchor through an xDS update – left previously issued session IDs valid, allowing a resumed session to be accepted without validation against the updated trust bundle. Configurations with a single CA produce byte-identical session IDs to the previous behavior; configurations with a multi-certificate bundle will issue new session IDs once after the upgrade, causing a one-time increase in full handshakes.
tls: Fixed a bug in TLS where a false positive IS_ENVOY_BUG assertion was triggered when a connection was torn down while asynchronous certificate selection was still in progress.
tls: Fixed a bug where OpenSSL was using glibc’s allocator instead of tcmalloc. This resulted in OpenSSL operating on a completely separate heap, defeating tcmalloc’s performance benefits on the TLS hot path and making all OpenSSL allocations invisible to tcmalloc heap profiling and memory dumps.
tls: Fixed a memory leak in the OpenSSL compatibility layer where SSL_get0_peer_certificates() called SSL_get_peer_certificate() without freeing the returned reference. Each call leaked one X509 refcount, preventing the certificate and its sub-allocations from being freed when the connection closed, causing unbounded memory growth in certain deployments.
tls: Fixed all code that accesses trusted CA certificates to return all trusted certificates instead of only the first one.
tls: Fixed upstream TLS client session caching so sessions are scoped by the effective SNI used for the connection. This prevents a session learned for one upstream SNI from being offered on a connection using a different SNI. The existing max_session_keys setting continues to limit the total number of cached sessions. This behavior can be temporarily reverted by setting runtime guard envoy.reloadable_features.scope_upstream_tls_session_cache_by_sni to false.
tls_inspector: Fixed a bug where JA4 fingerprint generation did not correctly encode non-alphanumeric ALPN characters as spec-compliant hexadecimal values. The fix is guarded by the runtime flag envoy.reloadable_features.ja4_alpn_hex_conversion_fix.
tracing: Fixed a crash during worker shutdown when the Datadog tracer is configured. A span still open when the worker destroys its thread local objects, for example the span of an in-flight mirror request, could flush its traces while the thread local cluster manager was being destroyed, and the request sent to the collector crashed in ClusterManagerImpl::getThreadLocalCluster() or used a destroyed collector cluster. The Datadog tracer now stops sending requests once its thread local tracer is destroyed, and looks the collector cluster up for every report.
udp: Fixed a bug where Envoy silently dropped zero-length UDP datagrams before issuing a socket operation. Empty datagrams are now sent through both connected and unconnected UDP sockets while preserving their packet boundaries. This behavior can be temporarily reverted by setting the runtime guard envoy.reloadable_features.udp_send_zero_length_datagrams to false.
udp: Fixed hot restart UDP packet forwarding for listeners bound with network_namespace_filepath. The draining parent now forwards the listener’s bind address and network namespace with each datagram and the child resolves the target listener from them, so listeners sharing an address across namespaces, and transparent listeners whose datagram destination differs from the bind address, receive their own forwarded datagrams instead of all being delivered to the first registered listener on that address.
upstream: Fixed a bug where upstream_bind_config with port 0 could cause ephemeral port exhaustion by reserving an ephemeral port during bind(). Envoy now automatically enables IP_BIND_ADDRESS_NO_PORT to defer port allocation until connect(). This change can be temporarily reverted by setting runtime guard envoy.reloadable_features.upstream_bind_config_fix_port_exhaustion to false.
upstream: Fixed a race condition affecting thread-aware load balancers (for example RING_HASH and MAGLEV) where, after a transient health-check failure followed by an immediate recovery, a worker thread could snapshot a stale load balancer factory and leave the recovered host absent from the ring/table until the next membership change. The thread-aware load balancer now rebuilds its factory state before the cluster manager posts the corresponding host update to the worker threads, so a worker can no longer snapshot a stale factory. When envoy.reloadable_features.enable_batch_aware_update is enabled (the default), the cluster manager accumulates per-priority host updates and posts them to the worker threads from the single end-of-cycle member-update callback (once for a whole batch host update, once after each individual update), instead of posting once per priority; mergeable health-check/weight/metadata updates still flow through the update merge window. The accumulated update is applied to each worker thread’s priority set as a single batch so the worker-local load balancer rebuilds once for the whole update instead of once per priority. This can be reverted by setting envoy.reloadable_features.enable_batch_aware_update to false. The thread-aware load balancer rebuilds its factory from the priority-update callback; when envoy.reloadable_features.coalesce_lb_rebuilds_on_batch_update is also enabled it instead defers the rebuild to the single end-of-cycle member-update callback, coalescing the per-priority rebuilds of a batch into one (which still lands before the batched post).
url_normalization: Fixes CVE-2026-73511

Strip path parameters from individual path segments per https://datatracker.ietf.org/doc/html/rfc3986#section-3.3

This behavioral change can be temporarily reverted by setting runtime guard envoy.reloadable_features.strip_path_parameters_per_segment to false.
url_normalization: Fixes CVE-2026-73551

Strip URL path parameters from dot and dotdot segments (segments that start with /..; or /.;). This allows path canonicalization to interpret them correctly. Stripping of path parameters from dot and dotdot segments occurs only if the normalize_path configuration option is enabled.

This behavioral change can be temporarily reverted by setting runtime guard envoy.reloadable_features.strip_dotdot_segments_with_parameters to false.
vhds: Fixed a bug where a VHDS subscription configured in an inline route_config was never started when its listener arrived over LDS after the server had finished initializing.
vhds: Fixed on-demand VHDS requests that could wait forever and leave the downstream request hanging: a repeated request for an already-answered alias waited for a response that never came, because the alias is already part of the delta subscription, so the xDS layer deduplicated the repeated interest and sent no new DeltaDiscoveryRequest at all (and even when a duplicate resource_names_subscribe entry was still sent on the wire, major control planes don’t implement the mandated re-send of an already-known resource). Requests queued against an update superseded while warming were also never resolved. Envoy now tracks which explicitly requested aliases the server has answered and resolves repeated requests for them locally from the published route configuration. Ids the server volunteered without a request are not answered locally, because no subscription guarantees further pushes for them. This behavior can be temporarily reverted by setting the runtime guard envoy.reloadable_features.vhds_answered_alias_cache to false.
wasm: Fixed a bug where Wasm plugins whose VM configurations differed could still share a single Wasm VM, and so silently run with the VM configuration of whichever plugin happened to be configured first. The runtime and the capability restrictions are now part of the VM identity, alongside the vm_id, the configuration, the code and the environment_variables, so plugins differing in either of them no longer share a VM.
zipkin: Fixed the Zipkin tracer to propagate the trace context with the B3 single b3 header when the downstream request uses it, instead of always injecting the multiple x-b3-* headers and forwarding the stale b3 header upstream. This behavior can be reverted by setting the runtime guard envoy.reloadable_features.zipkin_preserve_b3_single_header_format to false.

Removed config or runtime

Normally occurs at the end of the deprecation period

build: Removed Debian bullseye (11) packaging. Debian bullseye is end-of-life (LTS ended 31 August 2026) and its package repositories are no longer available on the main Debian mirrors. Bullseye .deb packages are no longer built or published.
build: Removed the envoy.network.connection_balance.dlb contrib extension (Intel DLB connection balancer), along with the dlb Bazel dependency, because the upstream Intel source archive is no longer available and there is no evidence of any users. See https://github.com/envoyproxy/envoy/issues/45491 for background.
ext_authz: Removed runtime flag envoy.reloadable_features.ext_authz_http_client_retries_respect_user_retry_on and legacy code paths.
ext_proc: Removed runtime flag envoy.reloadable_features.ext_proc_stream_close_optimization and legacy code paths.
generic_proxy: Removed the runtime guard envoy.reloadable_features.generic_proxy_codec_buffer_limit and the legacy code path it guarded. The generic proxy Dubbo, HTTP/1 and Kafka codecs now always fail decoding when the buffered data exceeds the connection buffer limit.
http: Removed the runtime guard envoy.reloadable_features.validate_upstream_headers and the legacy code path it guarded.
http2: Removed the runtime guard envoy.reloadable_features.safe_http2_options and the legacy code path it guarded. HTTP/2 connections now always fall back to the safe defaults (max concurrent streams of 1024, 16 MiB initial stream window and 24 MiB initial connection window) when the corresponding options are unset, and the unused legacy default constants are removed.
matching: Removed runtime flag envoy.reloadable_features.prefix_map_matcher_resume_after_subtree_miss and legacy code paths.
oauth2: Removed the runtime guard envoy.reloadable_features.oauth2_cleanup_cookies and the legacy code path it guarded. The OAuth2 filter now always removes the OAuth flow cookies (OauthHMAC, OauthExpires, RefreshToken, OauthNonce and CodeVerifier, including their suffixed names) from a request before it is forwarded upstream, so these cookies are no longer exposed to the backend service.
on_demand: Removed the runtime guard envoy.reloadable_features.odcds_over_ads_fix and the legacy code path it guarded. On-demand CDS over an ADS config source now always uses the per-cluster singleton subscription implementation (XdstpOdCdsApiImpl), the same mechanism used for xDS-TP based config sources.
on_demand: Removed the runtime guard envoy.reloadable_features.on_demand_track_end_stream and the legacy code path it guarded. The on-demand filter now always tracks the downstream end_stream state to decide whether a stream with a fully read body can be recreated, instead of rejecting all requests that carry a body.
original_dst: Removed the runtime guard envoy.reloadable_features.original_dst_rely_on_idle_timeout and the legacy code path it guarded. The original destination cluster now always checks whether hosts are in use by connection pools before removing them.
quic: Removed runtime flag envoy.reloadable_features.quic_fix_defer_logging_miss_for_half_closed_stream and legacy code paths.
quic: Removed runtime flag envoy.reloadable_features.quic_defer_logging_to_ack_listener and legacy code paths.
quic: Removed runtime flag envoy.reloadable_features.quic_signal_headers_only_to_http1_backend and legacy code paths.
quic: Removed runtime flag envoy.reloadable_features.quic_upstream_socket_use_address_cache_for_read and legacy code paths.
quic: Removed runtime flag envoy.reloadable_features.use_migration_in_quiche and legacy code paths.
quic: Removed runtime flag envoy.restart_features.validate_http3_pseudo_headers and legacy code paths.
rbac: Removed runtime flag envoy.reloadable_features.enable_cel_regex_precompilation and legacy code paths.
server: Removed runtime flag envoy.restart_features.raise_file_limits and legacy code paths.
tcp_proxy: Removed the runtime guard envoy.reloadable_features.tcp_proxy_odcds_over_ads_fix and the legacy code path it guarded.
tls: Removed runtime flag envoy.reloadable_features.reject_empty_trusted_ca_file and legacy code paths.
tracing: Removed the runtime guard envoy.reloadable_features.trace_refresh_after_route_refresh and the legacy code path it guarded. The HTTP connection manager now always refreshes the trace decision and decorator when the route is refreshed, and charges the tracing statistics from chargeStats rather than from the old un-refreshed code path.
upstream: Removed runtime flag envoy.reloadable_features.report_load_when_rq_active_is_non_zero and legacy code paths.
uri_template: Removed runtime flag envoy.reloadable_features.uri_template_match_on_asterisk and legacy code paths.
wasm: Removed the runtime guard envoy.reloadable_features.wasm_use_effective_ctx_for_foreign_functions and the legacy code path it guarded. The set_envoy_filter_state and clear_route_cache Wasm foreign functions now always resolve the effective context (contextOrEffectiveContext) instead of the current context.
websocket: Removed runtime flag envoy.reloadable_features.websocket_allow_4xx_5xx_through_filter_chain and legacy code paths.
websocket: Removed runtime flag envoy.reloadable_features.websocket_enable_timeout_on_upgrade_response and legacy code paths.

New features

access_log: Added the %LISTENER_NAME% access log command operator, which logs the name of the listener that accepted the downstream connection.
access_log: Added the DS_HS_BEG (downstream TLS handshake begin, i.e. when the ClientHello was received) and DS_HS_END (downstream TLS handshake end) time points to the %COMMON_DURATION% access log formatter. These are populated for both TLS and QUIC downstream connections. Also added the %DOWNSTREAM_CX_RTT% access log formatter returning the last measured round trip time of the downstream connection in milliseconds.
access_log: Added the DS_RX_HDR_END (downstream request headers fully received) time point to the %COMMON_DURATION% access log formatter.
admin: Added invert_filter query parameter to the /stats and /stats/prometheus admin endpoints. When set, the filter regex is inverted so matching stats are excluded from the output (e.g. /stats?filter=server&invert_filter).
ai_protocol_manager: Added envoy.http.ai_filters.transcoder (work in progress), an envoy.http.ai_filters extension that converts a declared AI endpoint’s request and response (unary and SSE streaming) payloads between a vendor’s schema and the canonical OpenAI Chat Completions intermediate representation (TO_IR and FROM_IR).
ai_protocol_manager: Added reserialize_body to the AI Protocol Manager filter. Set to DISABLE, it forwards the received request body byte for byte instead of re-serializing the parsed document, for AI filter chains that only read the request.
ai_protocol_manager: Added token_estimation to the envoy.http.ai_filters.request_info AI filter (alpha). When configured, the published envoy.data.ai.v3.RequestInfo record carries estimated_input_tokens, computed as ceil(tokens_per_byte * request payload bytes), for consumers that must budget before the provider reports usage.
ai_protocol_manager: Added an AI filter chain to the AI Protocol Manager filter (alpha): the filters field runs envoy.http.ai_filters extensions over a declared AI endpoint’s parsed payload before it is replayed. The first one, envoy.http.ai_filters.request_info, publishes request attributes as envoy.data.ai.v3.RequestInfo typed dynamic metadata before the request headers continue.
ai_protocol_manager: Added request-path statistics to the AI Protocol Manager filter, whose counters previously covered only the response path. request_parsed, request_parse_error and request_passthrough make the decode path’s outcomes visible. request_external_buffer_error and response_external_buffer_error count the 500 raised when the external buffer fails irrecoverably on each direction.
ai_protocol_manager: Added response-side LLM token-usage extraction to the AI Protocol Manager filter (alpha, work-in-progress API): streaming SSE and JSON responses in the OpenAI, Anthropic, and Gemini dialects are observed without stopping filter-chain iteration or mutating the response (extraction runs synchronously on the encode callbacks against a bounded side copy), and normalized usage is published as typed dynamic metadata (default namespace envoy.ai.token_usage): the authoritative record is envoy.data.ai.v3.TokenUsage, consumable via ext_proc typed metadata forwarding or any filter reading typed dynamic metadata. Inspection is scoped to routes carrying an AiProtocolManagerPerRoute configuration, with include_unconfigured_routes widening it to every route. Extraction behaves identically in the downstream and upstream (cluster, e.g. dynamic-forward-proxy egress) installations of the filter, and leaving request_handling unset yields a response-only installation whose request path is a pure passthrough.
ai_protocol_manager: Added the envoy.http.ai_filters.schema_validation AI filter to the AI Protocol Manager filter, which now runs request payload schema validation instead of the filter core. It validates against the request’s declared wire API, else a configured default, else one detected from the request, and rejects a violating payload with a 400 unless fail_open is set. A filter ahead of the AI Protocol Manager can name the request’s wire API with the envoy.ai.llm_protocol.request filter state object, which takes precedence over the route. The core’s request_schema_invalid counter is replaced by the AI filter’s own statistics.
ai_protocol_manager: The envoy.http.ai_filters.request_info AI filter (alpha) now stores the requested model as the envoy.ai.model.request filter state object, so filters, the router and access logs can read it as a string.
ai_protocol_manager: ai_protocol_manager: added usage_signal to TokenUsageExtraction. Setting usage_signal to SYNTHESIZE_TRAILERS adds empty response trailers at a clean end of stream when the response carries none of its own, allowing trailer-driven consumers such as ext_proc with response_trailer_mode: SEND to receive the published token usage metadata without streaming the response body.
alts: Added support for serverNameOverride() in ALTS transport sockets. Introduced ServerNameDecoratingTransportSocketOptions decorator.
aws: Added the aws_eventstream_parser filter. This filter extracts values from AWS EventStream HTTP response bodies (used by AWS Bedrock streaming APIs) and writes them to dynamic metadata for observability, logging, and cost tracking use cases.
build: Added --@envoy-docs//:build_sha and --@envoy-docs//:docs_tag build flags for passing documentation build metadata directly to the Sphinx runner.
compressor: Extended the compressor filter to support usage as an upstream HTTP filter. This can be used to apply request compression for OTLP traffic.
contrib: Added a new client_cert contrib HTTP filter (envoy.filters.http.client_cert) that forwards the downstream mTLS client certificate to upstreams using the Client-Cert and Client-Cert-Chain headers standardized by RFC 9440, sanitizing any incoming occurrences of those headers.
dns: Added case_insensitive to the DNS filter. When set, virtual domain names are matched case-insensitively while the response still echoes the client’s original query-name case. Defaults to false.
dynamic_modules: * dynamic_modules: added an HTTP filter ABI getter for the type URL and payload of typed
metadata, with C++, Go, and Rust SDK support.
dynamic_modules: Added envoy_dynamic_module_callback_get_runtime_bool, envoy_dynamic_module_callback_get_runtime_int and envoy_dynamic_module_callback_get_runtime_number so that dynamic modules can read boolean, integer and double values from the runtime. Each takes the key and the value to fall back to when the key is absent or holds a value of another type. They are exposed in the Rust SDK as get_runtime_bool, get_runtime_int and get_runtime_number, and in the Go and C++ SDKs as GetRuntimeBool / getRuntimeBool and friends on the new CommonHandle interface, which every config handle inherits.
dynamic_modules: Added envoy_dynamic_module_callback_http_has_filter_state ABI callback so a dynamic-module HTTP filter can check whether a filter state entry exists without reading or serializing the stored object. The C++, Go, and Rust SDKs expose the new callback.
dynamic_modules: Added envoy_dynamic_module_callback_http_set_dynamic_metadata_struct ABI callback that sets an entire dynamic metadata namespace from a serialized google.protobuf.Struct in one call, letting a module publish nested/structured metadata instead of only flat scalar keys. The Rust SDK exposes this as EnvoyHttpFilter::set_dynamic_metadata_struct, the C++ SDK as HttpFilterHandle::setMetadataStruct and the Go SDK as HttpFilterHandle.SetMetadataStruct.
dynamic_modules: Added envoy_dynamic_module_callback_http_set_dynamic_typed_metadata ABI callback that sets a typed dynamic metadata namespace from a serialized google.protobuf.Any. Unlike envoy_dynamic_module_callback_http_set_dynamic_metadata_struct it preserves the exact message type (via the Any type_url) in typed_filter_metadata, so consumers such as ext_authz (typed_metadata_context_namespaces) receive the original message rather than a lossy Struct. The Rust SDK exposes this as EnvoyHttpFilter::set_dynamic_typed_metadata, the C++ SDK as HttpFilterHandle::setTypedMetadata and the Go SDK as HttpFilterHandle.SetTypedMetadata.
dynamic_modules: Added a route specifier backed by a dynamic module. For each request the module keeps the route that route matching resolved, refines it, replaces it with one of the route templates it declares, drops it, or lets route matching carry on with the next route. It can be validated against the route table it replaces with shadow mode, which never changes the routing of a request.
dynamic_modules: Added a downstream wire byte accessor to the dynamic modules access logger ABI and Rust SDK, including support for locally generated responses with no upstream connection.
dynamic_modules: Added a dynamic modules HTTP/1 header formatter extension (envoy.http.stateful_header_formatters.dynamic_modules) that lets a dynamic module decide the casing of header keys written on the wire, which until now could only be done with the fixed policy of the preserve case formatter.
dynamic_modules: Added a dynamic modules cluster specifier extension (envoy.router.cluster_specifier_plugin.dynamic_modules) that lets a dynamic module select the upstream cluster for a request and replace the timeout, idle timeout, priority, request body buffer limit, cluster not found response code, hash policy, retry policy, metadata match criteria and request mirroring policies of the matched route. The selection context exposes the request headers, stream info attributes, dynamic metadata, the route name, the random value Envoy generated for cluster selection, and a routability query that reports host counts for a named cluster from the current worker, and the module is invoked again whenever a filter refreshes the route cluster. Custom counters, gauges and histograms can be defined during configuration and recorded during selection, and are emitted under the metrics_namespace prefix of DynamicModuleConfig. The Rust SDK exposes this through the cluster_specifier module and the cluster_specifier: arm of declare_all_init_functions!. See DynamicModuleClusterSpecifier for configuration details.
dynamic_modules: Added a dynamic modules early header mutation extension (envoy.http.early_header_mutation.dynamic_modules) that lets a dynamic module rewrite request headers before routing, tracing, request ID generation and any HTTP filter runs.
dynamic_modules: Added a dynamic modules string data input extension (envoy.matching.inputs.dynamic_module_string_data_input) that lets a dynamic module extract a string value from an HTTP request or response during match evaluation. The value is a standard string input, so map matchers such as an exact match map can dispatch on it, which lets a module select one of many matches with a single evaluation and without clearing the route cache. The Rust SDK exposes this through the matcher_data_input module and the declare_matcher_data_input! macro. See DynamicModuleDataInput for configuration details.
dynamic_modules: Added a route name setter to the dynamic modules route specifier. A module can now record the name of the route it produces through set_route_name on the route specifier context, so that a module built route carries an identity of its own for the %ROUTE_NAME% access log command operator and other route name consumers. The Rust SDK exposes this as RouteSpecifierContext::set_route_name.
dynamic_modules: Added generic secret subscriptions to the dynamic modules HTTP filter. During filter config initialization a module can subscribe to a generic secret by name, optionally passing a JSON serialized ConfigSource to fetch it over SDS, and read the current value per-stream or from the config context afterwards. Values are kept up-to-date as the SDS server pushes new versions. Available through the Rust, Go and C++ SDKs as subscribe_generic_secret/get_generic_secret, SubscribeGenericSecret/GetGenericSecret and subscribeGenericSecret/getGenericSecret respectively.
dynamic_modules: Added stats sink snapshot getters that expose each metric’s tag-extracted name and its tags (name/value pairs) for counters, gauges, text readouts, and histograms, so a dynamic module can reconstruct the dimensional metric names Envoy’s built-in formatters produce. Available through the Rust SDK MetricSnapshot tag accessors.
dynamic_modules: Added support for implementing a custom TLS handshaker as a dynamic module through the dynamic module TLS handshaker extension. A module can declare the handshaker capabilities, configure the SSL_CTX, and drive the TLS handshake. This works for both client and server TLS contexts.
dynamic_modules: Added the dynamic modules config validator (envoy.config.validators.dynamic_modules) for implementing xDS config validators with dynamic modules. This extends the dynamic modules ABI and adds Rust SDK support for config_validator registrations.
dynamic_modules: Added the envoy_dynamic_module_callback_bootstrap_extension_get_active_resource_names ABI getter so dynamic-module bootstrap extensions can enumerate the names of the config objects Envoy currently has active for a given resource kind. The kind (envoy_dynamic_module_type_bootstrap_active_resource_kind) is passed as an input and selects filter chains, clusters, transport socket matches or secrets, so future kinds are added without changing the ABI signature. The Rust SDK exposes this as active_resource_names. For the filter chain kind, an FCDS-delivered chain is reported only while an active listener’s matcher references it, so a name is returned only when the chain is both active and reachable, not merely committed in the process-wide FCDS manager. For the transport socket match kind, a name is reported only when it is present in every active cluster that has transport socket matches. The secret kind reports delivered dynamic TLS certificate, validation context, session ticket key and generic secrets.
dynamic_modules: Added the envoy_dynamic_module_callback_cluster_add_hosts_with_hostnames ABI callback so dynamic-module clusters can assign logical hostnames independently of socket addresses. Upstream TLS options such as auto_host_sni and auto_sni_san_validation can consume the logical hostname. Null or empty hostnames use the same synthesized hostname behavior as the existing callback. The Rust SDK exposes convenience and priority/locality-aware methods for the new callback.
dynamic_modules: Added the envoy_dynamic_module_callback_cluster_use_persistent_host_map cluster ABI callback and the Rust SDK EnvoyCluster::use_persistent_host_map method, which back the cross-priority host map of a dynamic modules cluster with a persistent map so that each host update costs O(log N) instead of a full copy of the map. The default flat map is unchanged.
dynamic_modules: Added the envoy_dynamic_module_callback_listener_filter_set_filter_state_typed and envoy_dynamic_module_callback_listener_filter_get_filter_state_typed ABI callbacks so a dynamic-module listener filter can write and read typed filter state, mirroring the existing bytes setter/getter. Unlike the bytes variant which stores a raw Router::StringAccessor, the typed setter uses the key’s registered ObjectFactory to build a properly typed filter state object, so a built-in Envoy filter that reads the key as a typed object can consume it. The Rust SDK exposes these as EnvoyListenerFilter::set_filter_state_typed and EnvoyListenerFilter::get_filter_state_typed.
dynamic_modules: Added the envoy_dynamic_module_callback_log_v2 callback, which logs a message and reports the module source location of the log statement instead of a location inside Envoy. The Rust, Go and C++ SDKs capture the call site automatically.
dynamic_modules: Added the envoy_dynamic_module_callback_network_filter_start_downstream_secure_transport ABI callback so a dynamic-module network filter can promote its downstream connection to TLS. The Rust SDK exposes this as EnvoyNetworkFilter::start_downstream_secure_transport.
dynamic_modules: Dynamic module clusters (envoy.clusters.dynamic_modules) can now use Envoy’s built-in load balancers. In addition to CLUSTER_PROVIDED, lb_policy may be set to LEAST_REQUEST, ROUND_ROBIN, RANDOM, RING_HASH, or MAGLEV; the module then supplies only host discovery and Envoy performs host selection.
dynamic_modules: Expose envoy_dynamic_module_callback_get_log_level in the Go HTTP filter and filter config handles as GetLogLevel and envoy_dynamic_module_callback_log_enabled as IsLogLevelEnabled.
dynamic_modules: Extended the dynamic modules Go SDK MetricSnapshot with histogram getters (HistogramCount, GetHistogram, HistogramBucketCount, and GetHistogramBucket) and per-metric tag getters (tag-extracted name, tag count, and individual tags) for counters, gauges, text readouts, and histograms, reaching parity with the Rust SDK so Go stats sinks can read histogram values and metric tags during flush.
dynamic_modules: The dynamic modules SDKs can now keep a tracing span past the event hook that created it. The Rust SDK get_active_span and spawn_child_span return owned spans, so a module can store a child span in the filter and finish it from a later event hook such as on_scheduled, which lets a span cover off-thread work. The C++ and Go SDKs already supported this and are now documented to guarantee it.
dynamic_modules: dynamic modules: HTTP filters and access loggers can now read timing markers for request receipt, downstream connection acceptance and close, TLS handshakes, upstream connects, the first response body byte received from upstream, and the final downstream ACK when available.
dynamic_modules: dynamic modules: added on_error to the dynamic modules input matcher to control the match result when a module cannot complete an evaluation, for example after a panic in the match hook. Defaults to NO_MATCH, preserving existing behavior. Set it to MATCH for deny-on-match trees where a missed match would otherwise let a request bypass the rule.
dynamic_modules: dynamic modules: added support for passing the value of xds.type.v3.TypedStruct to modules as JSON. This preserves a logical config type URL while keeping the payload readable in config dumps.
dynamic_modules: dynamic modules: added the connected upstream socket address and ordered host and connection ID attempt histories to the HTTP dynamic module ABI and C++, Go, and Rust SDKs.
dynamic_modules: dynamic modules: added the upstream HTTP protocol attribute to the dynamic module ABI and C++, Go, and Rust SDKs.
dynamic_modules: dynamic modules: added the upstream requested server name attribute to the dynamic module ABI and C++, Go, and Rust SDKs.
ext_authz: Added clear_route_cache_headers to the HTTP ext_authz filter. When set and clear_route_cache is true, an OK authorization response clears the route cache only when it sets, appends, or removes one of the listed request headers, or mutates a query parameter, instead of on any request mutation.
ext_authz: Added shadow_mode to the network ext_authz filter. When enabled, the filter calls the authorization service as normal but never closes the connection. The authorization decision is written to FilterState under envoy.filters.network.ext_authz so that a subsequent filter can read and optionally enforce it.
ext_authz: ext_authz: added emit_client_span and per-route emit_client_span to allow suppressing client-side egress spans for external authorization calls while preserving trace context propagation.
ext_proc: Added support for receiving typed dynamic metadata (typed_dynamic_metadata) from external processing servers.
ext_proc: ext_proc: added emit_client_span and per-route emit_client_span to allow suppressing client-side egress spans for external processing calls while preserving trace context propagation.
formatter: Exposed cel_config on the envoy.formatter.cel formatter configuration, allowing CEL runtime options such as string functions to be enabled for configured CEL formatters.
gcp_authn: Added audience to allow configuring the authentication service audience directly on the filter config, overriding cluster typed filter metadata.
gcp_authn: Added iam_access_token to support generating GCP IAM access tokens for target service accounts using substitution formatters.
gcp_authn: Added preserve_existing to support skipping token fetching and preserving the existing header intact when the target authorization header is already present.
gcp_authn: Added scopes and scopes to support configuring custom OAuth scopes for GCP access tokens and bound access tokens fetched from the metadata server.
gcp_authn: Added token_metadata_key to allow saving the fetched GCP authentication token to dynamic metadata under the filter config name namespace.
gcp_authn: Added statistics for the outcome of token fetches (token_fetch_success, token_fetch_failed), for token cache lookups (token_cache_hit, token_cache_miss) and for the request failures that are rejected with a local reply (iam_token_config_error, iam_token_resolution_failed, bound_token_fingerprint_unavailable). A failed token fetch previously produced only a log line, even though the request is forwarded upstream without a token. See statistics for the full list.
grpc_field_extraction: Added metadata_key to allow overriding the dynamic metadata key that an extracted field value is written to. If unset, the request field path is used, which is the previous behavior.
http: Added formatters to the header mutation filter to configure the formatter extensions used by filter and per-route mutation values, for example to enable CEL string functions.
http: Added min_concurrency_limit to the adaptive concurrency gradient controller to allow the minimum calculated concurrency limit to be configured separately from the concurrency used during minRTT recalculation.
http: Added a filter to limit the size of HTTP requests. See body size limit filter.
http: Added support for forwarding the issuer of the client certificate in the x-forwarded-client-cert (XFCC) header via the new issuer field of SetCurrentClientCertDetails. When enabled, the Issuer key is added in text format and the issuer field is added in JSON format. Defaults to disabled.
http: Added the %ROUTE_RESOLUTION_TIME_US% and %ROUTE_RESOLUTION_COUNT% access log command operators, reporting the total wall time in microseconds a stream spent resolving its route and how many times it resolved. Also added the record_route_resolution_stats option that records the downstream_rq_route_resolution_time_us and downstream_rq_route_resolutions histograms per stream when enabled.
http: The LocalResponsePolicy can now optionally preserve the existing response_code_details or set an explicit value via preserve_response_code_details (must be true if set) / response_code_details. Unset continues to clear details (legacy behavior).
http_compressor: * compressor: added support for content_type_matcher in CommonDirectionConfig, allowing fine-grained MIME type matching via string matchers.
jwt_authn: Added claim_path to claim_to_headers, which names the claim to copy as an explicit list of path segments instead of a single .-joined string. Each segment is matched in full, so claims whose own names contain dots are now addressable: the URL-namespaced claims issued by many OIDC providers, such as http://example.org/parent_token, and nested ones such as c.d inside a.b. Exactly one of claim_name and claim_path must be set; a claim_to_headers entry setting both or neither is now rejected at configuration load.
load_balancing: load_balancing: implemented the envoy.load_balancing_policies.load_aware_locality locality-picking load balancer. It weights localities by ORCA-derived utilization headroom and applies at all priority levels. ORCA data may be consumed in-band, out-of-band (setting enable_oob_load_report opens a per-host reporting stream, with optional connection overrides via oob_reporting_config), or both.
lua: Added filter_context to the Lua filter’s own configuration, so parameters shared by every route the filter serves no longer have to be repeated in each route’s LuaPerRoute. handle:filterContext() returns the route’s context when the route configures one and this one otherwise; a route’s context replaces rather than merges into it.
lua: Added package_paths and package_cpaths, and the same two fields on LuaPerRoute, which prepend module search patterns to a Lua VM’s package.path and package.cpath. This lets a script require modules from a configured location instead of only from the paths the interpreter searches by default. The patterns are applied before any configured code runs, including the run that validates the configuration, so a require at the top level of a script is resolved at config load time.
lua: Added shared_vm_id and LuaPerRoute.shared_vm_id to the Lua filter. Configurations that set the same id share one set of Lua VMs for every script whose contents and package search paths match, instead of each building its own. This is off by default: with the field unset the filter keeps building one set of VMs per configured script.
lua: Added requestHeaders() to the Lua HTTP filter’s stream handle, allowing a script to read the request headers from envoy_on_response as well as envoy_on_request. Returns nil if the stream has no request headers.
lua: Added an optional ns parameter to the metadata() methods of the HTTP Lua filter’s stream handle and route object. It lets a script read the route metadata under any namespace rather than only the one named by the filter config name.
lua: Added the base64Decode() method to the HTTP Lua filter’s stream handle, the inverse of the existing base64Escape(). It returns nil when the input is not valid base64.
mcp: Added early_terminate_when_routable to the MCP filter. When enabled, the filter stops parsing and buffering the request body as soon as all required routing attributes for the request’s method have been collected (for example method and params.name for tools/call), instead of buffering up to max_request_body_size and waiting for the root JSON object to close. This decouples routing from body size so a large trailing payload (for example params.arguments) does not need to be buffered just to route the request. Early termination is skipped when reject_duplicate_keys, trace context or baggage propagation is configured, or when attribute_source is not BODY. Defaults to false.
mcp: Added attribute_source support for extracting MCP request attributes from the request body, verifying request headers against the body, or using Mcp-Method and Mcp-Name headers for a body-free fast path.
mcp: Added max_supported_protocol_version support for limiting requests to known MCP protocol revisions up to a configured maximum, including 2026-07-28-specific header and transport semantics, protocol-version consistency checks between MCP-Protocol-Version and request metadata, and Base64 decoding for encoded Mcp-Name values. Header mismatch and unsupported-version errors are returned as standardized JSON-RPC error responses.
mcp: Added a NOOP traffic mode to McpFilter.
mcp: Added support for server/discover, subscriptions/listen, and tasks/* method groups, and support for extracting params.taskId in MCP JSON parser.
mcp: Added validation to the MCP filter that MCP 2026-07-28 requests include io.modelcontextprotocol/clientCapabilities in params._meta when traffic_mode is REJECT_NO_MCP. Requests missing the field are rejected with HTTP 400 and JSON-RPC error -32602. Only JSON-RPC requests are checked; notifications are exempt. In this mode, attribute_source: HEADERS now parses the request body for 2026-07-28 requests.
mcp: McpFilter now allows per-route config for clear_route_cache, parser_config, request_storage_mode, and reject_duplicate_keys.
mcp_json_rest_bridge: Added resultType: "complete" to transcoded tools/call and locally generated tools/list results for MCP 2026-07-28 stateless requests.
mcp_json_rest_bridge: Added support for the MCP 2026-07-28 stateless lifecycle. The bridge now handles server/discover locally, advertises all supported MCP protocol versions, and supports configurable discovery cache metadata through server_info.server_discovery_cache_config. Stateless requests reject the legacy initialize and notifications/initialized handshake, while earlier protocol requests continue to use the existing initialization flow.
mcp_json_rest_bridge: Added validation of Mcp-Method and Mcp-Name request headers for MCP 2026-07-28, including Base64-decoded Mcp-Name values. Missing, malformed, or mismatched headers are rejected with a HeaderMismatch error.
mcp_json_rest_bridge: Added validation of the MCP-Protocol-Version request header for stateless MCP 2026-07-28 requests. A missing, duplicated, or mismatched header (compared with params._meta["io.modelcontextprotocol/protocolVersion"]) is rejected with a 400 HeaderMismatch (-32020) JSON-RPC error. A protocol version that is not supported for stateless requests is rejected with a 400 UnsupportedProtocolVersion (-32022) JSON-RPC error that lists the requested and supported versions.
mcp_transcoder: Added per_route_only config to McpJsonRestBridge. When set, the filter will take no action unless per-route configuration is available.
network_ext_proc: Added support for evaluating CEL connection attributes and filter state in network external processor. Configured via connection_attributes and sent in ProcessingRequest.attributes.
oauth2: Added TLS_CLIENT_AUTH to the OAuth2 credential injector, which authenticates the client to the token endpoint using mutual TLS, implementing OAuth 2.0 Mutual-TLS Client Authentication as defined in RFC 8705. The token request body contains only grant_type and client_id, and the client certificate is taken from the transport socket configured on the token_endpoint cluster. client_secret is not required when this auth type is used.
oauth2: Added assertion_audience to the OAuth2 filter’s private key JWT configuration, setting the aud claim of the client assertion (for example to the authorization server’s issuer identifier, which some identity providers require). If unset, the aud claim remains the configured token_endpoint URI. The token_secret may now also be supplied as a multi-entry generic secret with private_key and key_id entries; when a key_id entry is present, its value is set as the kid header parameter of the client assertion and is rotated together with the signing key. Existing single-value secrets are unaffected and emit no kid header.
oauth2: The OAuth2 filter now includes a RequestId tag in its application log lines, matching the access log %STREAM_ID% / x-request-id value, so operators can correlate OAuth2 application logs with access logs on a per-request basis.
opentelemetry: Added support for resource_detectors to the OpenTelemetry access logger.
overload_manager: Overload manager: added support for multiple envoy.overload_actions.reduce_timeouts actions with configured instance names, allowing different resource monitors to drive timer groups independently.
quic: Added support for P-384 and P-521 ECDSA leaf certificates for QUIC downstream connections, in addition to the previously supported P-256. This can be temporarily reverted by setting the runtime guard envoy.reloadable_features.quic_support_additional_ecdsa_curves to false.
quic: Added support for memory optimization in QUIC by resetting the internal SSL object after the handshake finishes. This can be enabled by setting the runtime guard envoy.reloadable_features.quic_enable_reset_ssl_after_handshake to true.
quic: QUIC downstream listeners now support client certificate authentication (mutual TLS). When a filter chain’s downstream TLS context sets require_client_certificate, the server requires a client certificate during the handshake. When the filter chain instead configures a certificate validation context without require_client_certificate, the server requests but does not require one (optional mutual TLS), and the handshake still succeeds if the client presents no certificate. Whenever the client presents a certificate, its fields are exposed to consumers such as x-forwarded-client-cert, RBAC, and access logs, and it is validated against the trust anchor of the filter chain matched for the connection unless trust_chain_verification is ACCEPT_UNTRUSTED. A filter chain that requires a client certificate must also configure a validation context that can act on the presented chain, that is one with a trusted_ca, a custom_validator_config, or trust_chain_verification set to ACCEPT_UNTRUSTED. The validation context may be delivered over SDS, in which case handshakes are rejected until the secret arrives. Session resumption and early data default to off on filter chains that configure an inline validation context, because QUIC does not re-validate the client certificate when a session is resumed.

This behavior can be reverted by setting the runtime guard envoy.reloadable_features.quic_mtls_server_enabled to false, which restores the previous behavior of rejecting QUIC listeners that require a client certificate and not requesting a client certificate for filter chains that configure an optional validation context. The session resumption and early data default can be reverted separately by setting envoy.reloadable_features.quic_mtls_resumption_disabled_by_default to false.
quic: Upstream QUIC connections now present the client certificate configured in the cluster’s upstream TLS context when the upstream server requests one. Previously configured client certificates were silently not sent over HTTP/3. Client certificates using a private key provider are not supported over QUIC and are now rejected at configuration load time. This behavior change can be reverted by setting the runtime guard envoy.reloadable_features.quic_upstream_client_certificates to false; the guard is evaluated when a cluster’s transport socket is created, so flipping it takes effect on clusters created or updated afterwards.
ratelimit: Added unit_multiplier to the Rate Limit Service descriptor override and response APIs to support rate limit periods that span multiple units of time. The HTTP rate limit filter uses the response value when calculating the window in X-RateLimit-Limit headers.
ratelimit: Added an opt-in enable_retry_after_header option to the global rate limit filter and an equivalent local rate limit option. For enforced 429 responses, enabling the option allows the filter to add a Retry-After header. For the global rate limit filter, the rate limit service response must contain at least one OVER_LIMIT descriptor status; the header value is the largest duration_until_reset among those statuses. For the local rate limit filter, the header value is the time until the token bucket that rejected the request has a token available. Both values are expressed in seconds and clamped to at least 1. Both filters preserve an existing Retry-After header. The option is disabled by default and has no effect on responses with any status code other than 429.
ratelimit: Extended the global rate limit filter so a hits_addend format string may reference response headers via %RESP()% on the apply_on_stream_done path, resolving the value from the upstream response without a dynamic-metadata hop. On the request path the response is not yet available, so %RESP()% resolves empty and the descriptor is dropped, as before.
ratelimit_descriptors: Added a new envoy.rate_limit_descriptors.jwt_claim rate limit descriptor extension that extracts a named claim from a JWT found in an HTTP header or cookie and uses it as a descriptor value. This is useful when JWT validation is performed elsewhere (e.g. by the application, or by an upstream mTLS-authenticated service) and Envoy only needs to rate limit based on the claim value. Note that this extension does not verify the JWT signature.
redis_proxy: Added RESP3 protocol support to the Redis proxy via the new protocol_version listener setting (default RESP2 keeps the existing behavior). When set to RESP3, downstream clients negotiate with an explicit HELLO 3 handshake — data commands sent beforehand are rejected with -NOPROTO and counted by the new downstream_rq_noproto counter — and every new upstream connection performs a HELLO 3 handshake (combined with AUTH or AWS IAM credentials when configured, followed by READONLY where applicable) before serving traffic; requests issued during the handshake are held and replayed in order, and negotiation failures are tracked by the new per-cluster upstream_resp3_hello_failure counter. Independently of the setting, the proxy now answers HELLO, CLIENT SETNAME and CLIENT SETINFO locally so that modern Redis clients can complete their connection setup, and the codec understands all RESP3 frame types, down-converting them for RESP2 connections.
redis_proxy: Added support for proxying the CLUSTER SHARDS command. Like the other supported CLUSTER introspection subcommands (INFO, SLOTS, KEYSLOT, NODES), it is forwarded to a single random upstream shard and the reply is returned to the client unmodified.
reverse_tunnel: Added experimental JWT authentication for the reverse tunnel handshake via the new jwt_validator field on the envoy.filters.network.reverse_tunnel filter. When configured, the bearer token carried in the handshake request is verified (signature, issuer, audiences, and exp) before the connection is accepted and its socket registered, so a forged or expired token cannot establish a usable reverse tunnel. A jwt_validator block requires an issuer, and tokens without an exp claim are rejected. Verified claims are published as dynamic metadata so the existing validation block can bind a claimed identifier to a verified claim via %DYNAMIC_METADATA(namespace:claim)%. The JWKS may be supplied inline via local_jwks or fetched over HTTP via remote_jwks; remote keys are fetched and refreshed in the background (at startup and every cache_duration) so handshake verification stays synchronous.
reverse_tunnel: The downstream reverse-tunnel initiator (envoy.bootstrap.reverse_tunnel.downstream_socket_interface) now accepts maintain_interval to control how often each host is re-checked and missing tunnels are dialed. Unset keeps the historical 10s default. The existing 15% upward jitter still applies. The minimum allowed value is 100ms.
reverse_tunnel: The downstream reverse-tunnel initiator (envoy.bootstrap.reverse_tunnel.downstream_socket_interface) now includes two additional identifiers in the HTTP handshake it sends to the acceptor: x-envoy-reverse-tunnel-worker-id (the initiator worker dispatcher name, e.g. worker_2) and x-envoy-reverse-tunnel-connection-id (the initiator’s per-connection id). Both are surfaced in the initiator access log via the new worker_id and connection_id fields of the envoy.reverse_tunnel.initiator dynamic metadata namespace. The upstream acceptor (envoy.bootstrap.reverse_tunnel.upstream_socket_interface) now parses these headers and exposes them on every reverse-tunnel lifecycle event as the initiator_worker_id and initiator_connection_id fields of the envoy.reverse_tunnel.lifecycle dynamic metadata namespace and as the envoy.reverse_tunnel.initiator_worker_id / envoy.reverse_tunnel.initiator_connection_id connection filter-state keys. Together these let tunnels originating from different workers/connections of the same initiator be told apart and correlated across both ends.
router: Added filter_state, an internal redirect predicate that gates redirect decisions on a boolean filter-state object set earlier in the request lifecycle (for example by a Lua filter, ext_proc, set_filter_state, or a dynamic module). The predicate follows the redirect when the boolean value is true, enabling per-request redirect control without changing route matching.
router: Added route_specifiers, a new extension point that post-processes the route resolved by route matching. Specifiers are executed in order and the output of each is the input of the next, and may be configured on a route configuration, on a virtual host and on a route, evaluated in that order. A specifier may refine the route it is given, drop it, or generate a route of its own for a request that matched none, so routing decisions can be customized without changing route matching.
router: Added a new priority group cluster specifier that splits the candidate clusters of a route into a list of named groups. The group is selected based on the attempt count of the request (the initial attempt uses the first group, the first retry uses the second group, and so on, staying on the last group once the attempts go past the end of the list) and the target cluster is then selected from the group based on the cluster weights. The groups can be overridden per request by an optional dynamic metadata namespace. This must be used together with refresh_cluster_on_retry to ensure the target cluster is re-selected on every attempt and the request is retried in a different priority group.
sds: Added poll_interval to filesystem configuration sources. When used with SDS, Envoy polls both the SDS configuration and the same secret files that are watched in event-based mode, allowing rotation when filesystem notifications are unreliable or a custom deployment model does not generate the move or modification events handled by watching the path or directory.
sockets: Added a new validate_network_namespaces option to BindConfig. When set, the network_namespace_filepath of every source address in the bind config is validated at configuration load time, and the configuration is rejected if a referenced Linux network namespace cannot be opened.
stat_sinks: Added scale_histogram_units_to_milliseconds to the statsd, DogStatsD and Graphite statsd sinks. When enabled, histogram samples are scaled to milliseconds according to the histogram’s unit before being reported as timers: samples of histograms recording microseconds are divided by 1000 and reported as a fractional millisecond value, while histograms recording milliseconds or without a declared unit are reported unchanged. By default every sample is still reported unchanged with an ms suffix regardless of its unit.
stats: Added Nanoseconds to the histogram units, so histograms created by extensions such as the Lua filter (unit "nanoseconds") and by embedders can record nanosecond samples. Timespans flushing to such a histogram record the elapsed nanoseconds, and the statsd, DogStatsD and Graphite statsd sinks scale the samples to milliseconds when scale_histogram_units_to_milliseconds is enabled. No Envoy histogram uses the new unit, so existing output is unchanged. Note that the default Prometheus and OpenTelemetry bucket boundaries assume milliseconds, so nanosecond histograms need explicit histogram bucket settings.
stats: Added per-cluster and per-listener stats_matcher configuration that overrides the bootstrap stats_config matcher for the specific cluster or listener. When this field is configured, legacy envoy.stats_matcher metadata is ignored.
tcp_proxy: Added load shed point envoy.load_shed_points.tcp_proxy_upstream_connect and metric downstream_cx_overload_close to shed downstream connections when establishing upstream connections under resource pressure.
tcp_proxy: Added load shed point envoy.load_shed_points.tcp_proxy_on_data to close downstream TCP connections when receiving data under resource pressure, and added the downstream_cx_overload_close counter stat to the TCP proxy filter to track connections closed by load shedding.
tcp_proxy: Added propagation of downstream TCP RST to upstream for direct TCP proxy connections on Linux when the detected close type is RemoteReset. This behavioral change can be temporarily reverted by setting runtime guard envoy.reloadable_features.propagate_downstream_rst_to_upstream to false.
tls: Added support for per-certificate TLS parameter overrides via the tls_params field on TlsCertificate. When set on a server certificate, any specified fields override the corresponding context-level TLS parameters for that certificate during the TLS handshake; unset fields continue to use the context-level values. This allows different cipher suites, ECDH curves, protocol versions, signature algorithms, and compliance policies per certificate. These parameters do not affect certificate selection, which remains based on SNI and the client’s ECDSA and OCSP capability, and are applied after a certificate has been selected. With multiple certificates, each certificate’s parameters must be compatible with the clients that select it, since an incompatible selected certificate fails the handshake rather than falling back to another certificate. This field is not supported on client certificates and is ignored. It also has no effect on QUIC/HTTP3 downstream connections.
tls: Allow multiple tls_certificates in a client context, for CommonTlsContext , when a custom_tls_certificate_selector is explicitly defined with max_sesion_keys set to 0.
tls: The SPIFFE certificate validator now supports additional verification of the upstream peer certificate SAN names via the well-known filter state envoy.network.upstream_subject_alt_names. Both the overridden SAN list and the configured SAN matchers must match if both are present. This behavior change can be reverted by setting the runtime guard envoy.reloadable_features.spiffe_validator_use_upstream_subject_alt_names to false.
tracing: Added a value_type field to tracing custom tags. When set to INT, DOUBLE or BOOL, tracers that support typed span attributes (such as OpenTelemetry) emit the custom tag as a native integer, floating-point or boolean span attribute instead of a string. Applies to every custom tag type. When unset (UNSPECIFIED) or set to STRING the tag keeps the preexisting string behavior for all tracers.
tracing: tracing: added set_instrumentation_scope option to the OpenTelemetry tracer to allow controlling the emission of the instrumentation scope name and version in traces.
tracing: tracing: added an exporter extension point to the OpenTelemetry tracer configuration, allowing the use of custom tracing exporters.
udp: Added formatters to the UDP proxy tunneling config, allowing formatter extensions to be used in the substitution formatters of headers_to_add.
udp: Added UDP session handoff during hot restart: the parent instance keeps serving established UDP sessions while draining and forwards datagrams of unknown sessions to the child instance over the hot restart RPC. Connectionless UDP listeners no longer stop functioning between drain start and parent shutdown. The set of served sessions is the set a listener filter has registered. Currently only udp_proxy registers its sticky sessions, which the parent keeps serving until they reach the idle timeout. Forwarding to the child can be disabled with the envoy.reloadable_features.udp_hot_restart_session_handoff runtime guard, in which case the draining parent keeps serving all datagrams itself until it exits.
udp: udp_proxy: native UDP upstream sockets now honor the cluster or bootstrap upstream_bind_config, including the selected source address and pre-bind socket options. Configure source port 0 to let the kernel allocate an ephemeral port for each UDP proxy session.
upstream: Added preconnect_enabled_metadata to restrict upstream preconnects to hosts whose endpoint metadata matches the configured matcher. Non-matching hosts receive connections only for on-demand requests. Suppressed preconnects increment a new upstream_cx_preconnect_skipped counter.
upstream: Added a new per_worker_subset load balancing policy. Each Envoy worker maintains its own subset of upstream hosts and load-balances requests across only that subset, decoupling per-worker connection-pool size from total cluster size. Useful for large upstream clusters where the upstream enforces a short server-side HTTP keepalive timeout. Two partitioning strategies (EQUAL_PARTITIONS and RANDOM_PARTITIONS) can be combined with three within-subset host-selection strategies (SIMPLE_ROUND_ROBIN, ENVOY_ROUND_ROBIN delegating to the stock RoundRobin LB, and ENVOY_P2C delegating to the stock LeastRequest LB). Per-worker fallback eliminates synchronized cluster-wide connection-pool churn that a cluster-wide healthy-fraction check would produce.
upstream: Added support for configuring multiple health checkers on a single cluster. When multiple health checks are specified, a host is considered healthy only when all checkers report it as healthy. Each health check entry requires a unique name field for per-checker stats tracking.
upstream: Added the UDP health checker, which sends a configured UDP datagram and waits for an exact response payload.
watchdog: Added envoy.watchdog.backtrace_action, a new watchdog action that logs a stack backtrace of stuck threads when the watchdog fires. A configurable cooldown prevents duplicate backtraces for the same thread.

Deprecated

config: The ConfigValidatorFactory::typeUrl() method is deprecated in favor of ConfigValidator::typeUrl(), which resolves the xDS type url from the validator instance and can therefore depend on the validator’s configuration. This only affects extension code, not configuration. Both methods are now non-pure and default to an empty value. Envoy calls ConfigValidator::typeUrl() first and falls back to the deprecated ConfigValidatorFactory::typeUrl() only when it returns empty, so a validator only needs to implement the new instance method. The deprecated method keeps working and will be removed once the in-tree and out-of-tree extensions have migrated. Envoy itself builds with -Wno-deprecated-declarations, so this deprecation is only visible to out-of-tree builds that enable the warning.
dynamic_modules: The envoy_dynamic_module_callback_log callback is deprecated in favor of envoy_dynamic_module_callback_log_v2, which additionally reports the module source location of the log statement. The deprecated callback keeps working and will be removed after the deprecation period in the ABI compatibility policy.
dynamic_modules: dynamic modules: deprecated envoy_dynamic_module_callback_access_logger_get_timing_info. Use envoy_dynamic_module_callback_access_logger_get_timing_info_v2 for the expanded timing snapshot.
http: The HTTP filter factory base classes FactoryBase, ExceptionFreeFactoryBase, and DualFactoryBase, together with the createFilterFactoryFromProto() entry points on NamedHttpFilterConfigFactory and UpstreamHttpFilterConfigFactory, are deprecated in favor of UnifiedFactoryBase and its single createHttpFilterFactoryFromProtoTyped() entry point, which serves both the downstream and the upstream HTTP filter chains. This only affects extension code, not configuration. createFilterFactoryFromProto() is no longer pure virtual: it now defaults to delegating to createHttpFilterFactoryFromProto(), so a factory that implements the interfaces directly only needs to implement the new entry point. The deprecated classes and methods keep working and will be removed once the in-tree and out-of-tree extensions have migrated. Note that Envoy itself builds with -Wno-deprecated-declarations, so these deprecations are only visible to out-of-tree builds that enable the warning; such builds can pass -Wno-deprecated-declarations to keep compiling while the migration is in progress.
reverse_tunnel: Deprecated auto_close_connections in the reverse tunnel network filter. The handshake connection is now always detached once the acceptance response is flushed, so the field is ignored.
wasm: The PluginConfig.capability_restriction_config field is deprecated in favor of the new VmConfig.capability_restriction_config field. The restrictions are applied when the Wasm VM is created and are shared by every plugin running in that VM, so they are a property of the VM rather than of an individual plugin. The deprecated field keeps working: when it is set and the VM level field is not, it is used to populate the VM level one.