Consul
$
npx @buildinternet/releases get consul-releasesMon
Wed
Fri
OctNovDecJanFebMarAprMayJunJulAugSepOct
LessMore
Releases4Avg1/moVersionsv2.0.2 to v2.0.4
Last Checked
11h ago
npx @buildinternet/releases get consul-releasesBREAKING CHANGES:
consul connect redirect-traffic command from iptables/ip6tables to nftables (nft). The command now uses the sdk/nftables package, which applies all IPv4 and IPv6 rules atomically in a single pass via the nftables inet address family, instead of separate iptables and ip6tables invocations. [GH-23785]iptables previously also accepted user names, UID ranges, and host names. [GH-23785]consul connect redirect-traffic does not remove iptables/ip6tables rules created by earlier Consul versions. This only matters when upgrading an existing host or VM in place. Before running the new command on an existing host or VM, remove any Consul-managed iptables rules first (see the transparent proxy upgrade guide for a reference cleanup script). Leaving both rule sets in place can cause double NAT or inconsistent traffic redirection. [GH-23785]nftables instead of iptables. Scripts that run iptables inside the Consul image must be updated. [GH-23785]iptables package with a new nftables package for transparent proxy traffic redirection. Traffic redirection now requires the nft binary and a Linux kernel with stateful NAT support in nftables inet family chains, available upstream starting with Linux kernel version 5.2 (some distributions backport this support onto an earlier nominal kernel version, for example RHEL 8+, kernel 4.18+, and its derivatives). There is no automatic fallback to iptables/ip6tables, so hosts without this kernel support will fail to set up transparent proxy redirection. [GH-23785]SECURITY:
brace-expansion to address GHSA-rgw5-rvv9-x895 (DoS via unbounded intermediate arrays). [GH-23786]fast-uri to address GHSA-7p8r-x3mc-p8w7 (Host Confusion via backslash authority introducer). [GH-23786]socket.io-parser to address CVE-2026-69185 (Zero-attachment Memory Exhaustion). [GH-23786]FEATURES:
ai block to service definitions, letting a workload declare its AI role (inference-model, mcp-server, or ai-agent) and role-specific configuration in a single registration via config files and HTTP APIs. [GH-13032]http2 and grpc as valid API Gateway listener protocols, enabling end-to-end HTTP/2 and gRPC from client to backend. Previously only http and tcp were accepted; the gateway would silently downgrade gRPC/HTTP2 traffic to HTTP/1.1 on the upstream hop. The new protocols render the correct downstream HTTP connection manager (HTTP/2 codec, h2,http/1.1 ALPN) and align the upstream cluster to HTTP/2, closing the parity gap with Ingress Gateway. [GH-23784]<service>.api-gateway.consul. Populates the gateway-services catalog mapping for API Gateways and adds DNS SANs to the API Gateway Connect leaf certificate, bringing API Gateway to parity with Ingress Gateway. The feature is version-gated and backfills pre-existing gateways automatically once all servers are upgraded. [GH-23647]TLS { Enabled = true } is set on an api-gateway config entry with no inline/file-system/SDS certificate, the gateway terminates downstream TLS using its Connect CA leaf certificate (with *.api-gateway.consul DNS SANs). Explicitly configured certificates continue to take precedence and are served as SNI overrides alongside the leaf default chain. [GH-23647]PassiveHealthCheck to gateway-level default upstream limits and route service-level limit overrides, so passive health checking (Envoy outlier detection) can be configured for services routed through an API gateway. [GH-23783]consul connect envoy -gateway=inference launches the policy processor alongside Envoy and enforces a minimum Envoy version for the inference gateway. The ext_proc socket is derived per proxy instance; -ext-proc-socket-dir overrides its directory (default /var/run/consul).ai.role of ai-agent or mcp-server, including dynamic client registration with IBM Security Verify, key rotation, JWKS publication, and SDS delivery of the OAuth client credential to ai-agent Envoy sidecars.ai.role of ai-agent or mcp-server with IBM Security Verify using private_key_jwt, persist DCR/JWKS in KV, and deliver the client credential to Envoy as SDS GenericSecret JSON for ai-agent workloads. Cluster IAM is an OIDC ACL auth method with EnableServiceDCR=true (agent oauth.iam HCL is rejected).ecdh_curves to mesh config entries (MeshDirectionalTLSConfig) and tls_ecdh_curves to agent TLS configuration (TLSProtocolConfig) to support configuring allowed ECDH/KEM curves for sidecar mTLS and agent RPC/HTTPS/gRPC, with automatic hybrid Post-Quantum X25519MLKEM768 curve injection when tls_min_version is set to TLSv1_3. [GH-23884]feature_gates { bootstrap = { ... } } server configuration block seeds the first Raft policy. The leader continuously reconciles desired policy, minimum-version eligibility, and registry defaults into a committed resolved status. Every server runs an atomic cache populated only from committed FSM state; agents consume an indexed resolved snapshot and fail closed until the first generation is delivered. New operator CLI commands consul operator feature list, get, and set provide CAS-safe, ACL-protected inspection and control of feature gates.
feature-gates: Gate API Gateway HTTPRoute upstream-routing composition (PR #23294) behind the experimental feature api-gateway-upstream-routing. When disabled (default), the legacy HTTPRoute-to-router shape is preserved exactly. When enabled by an operator after a full datacenter upgrade, service-router rules are composed with HTTPRoute matches, resolver subset definitions are preserved, and discovery-chain watches use the correct HTTP protocol override. Existing agentless API Gateway proxy states are rebuilt automatically when the effective gate value changes.
operator: Add GET /v1/operator/features, GET /v1/operator/feature/:name, and PUT /v1/operator/feature/:name HTTP endpoints for reading and updating feature gate state. Reads require operator:read; writes require operator:write. All reads are blocking-query capable. [GH-23791]inference-gateway service kind and config entry. The gateway terminates inbound mesh mTLS, enforces intentions, and applies PII and routing policy through an ext_proc filter to a co-located policy processor. Model upstreams are services registered with ai.role = "inference-model" that intentions allow and that are reachable through the mesh (connect-enabled or behind a terminating gateway). The gateway connects to them over mTLS and selects them by capability, failing over across providers by priority.consul.usage.mesh_gateway_remote.count and consul.usage.mesh_gateway_local.count report the number of service instances configured to use each mesh gateway mode, and consul.usage.mesh_gateway_none.count reports the number of service instances with no explicit mesh gateway mode (unset or none).consul.usage.mesh_gateway_enabled, which reports whether any mesh gateway instance is registered in the cluster. It is independent of the mesh gateway mode counts: those count connect-proxy sidecars by their upstream routing mode, while this reports whether the gateway capability is deployed at all.TerminatingGateway config entry now carries a CredentialInjection block (ext_proc processor UDS path and message timeout) and each linked service can declare a Credential binding (Mode: "inject" with a BindingID). At egress, an ext_proc filter calls a co-located processor that injects the bound credential (e.g. a provider API key sourced from Vault) into the upstream request, so credentials are never stored in mesh config or exposed to clients.ConfigFetcher interface (and agent implementation) to expose DNS recursors. [GH-23698]IMPROVEMENTS:
invert field to HTTPHeaderMatch in http-route config entries, enabling negated header match conditions (e.g. "route when header is absent" or "route when header value does NOT match"). Brings API Gateway to parity with the Ingress Gateway's existing Invert support in service-router. [GH-23817]consul.usage.client_agent_metric_included reports whether client-agent metrics collection is enabled. consul.usage.client_auto_encrypt_enabled.count and consul.usage.server_auto_encrypt_enabled.count reports the number of clients and servers, respectively, with auto_encrypt.tls enabled. consul.usage.auto_encrypt_enabled combines both counts into a single boolean — true if auto-encrypt is enabled anywhere in the cluster be it client or server.consul.usage.api_gateway.enabled reports whether any API gateway instances are currently registered in the cluster, consul.usage.mesh.service.percentage reports the percentage of service instances that are connect-proxy sidecars and consul.usage.transparent_proxy.enabled.percentage reports the percentage of connect-proxy instances that have transparent proxy mode enabled.builtin/ext-proc Envoy extension now supports Unix domain socket processor targets via Target.Path.BUG FIXES:
x-forwarded-client-cert (XFCC) header (such as auto-registered DNS SANs) are now properly accepted by the RBAC principal regular expression. [GH-23920]consul intention create now creates and replaces intentions via the exact (config-entry) intention API used by the UI, which supports non-default admin partitions (for example -partition team-a). Per-source metadata (-meta) continues to use the deprecated legacy intention API and is only supported in the default partition; combining -meta with a non-default partition now fails with a clear, actionable error. To attach metadata to intentions in an admin partition, use a service-intentions config entry with entry-level Meta via consul config write.ai-agent / mcp-server). Explicit HTTP upstreams to inference-gateway or inference-model destinations no longer get consul-obo-outbound and fail closed with "OBO audience not configured".default_intention_policy for three server-side surfaces — the intention check API (consul intention check, GET /v1/connect/intentions/check), the service topology view (Internal.ServiceTopology), and transparent proxy upstream discovery (Internal.IntentionUpstreams) — and fell back to acl.default_policy instead. This only changes behavior when default_intention_policy and acl.default_policy are set to opposite values; when they agree (or default_intention_policy is unset) there is no change. With default_intention_policy=deny and acl.default_policy=allow, the check API and topology previously reported connections as allowed even though they were denied; with default_intention_policy=allow and acl.default_policy=deny, transparent proxy sidecars under-provisioned their upstreams and could fail to reach services they were authorized to reach. Envoy RBAC enforcement was not affected and already honored default_intention_policy.BREAKING CHANGES:
+ent.fips1402 to +ent.fips1403, and package and container artifacts change from the F2 suffix to F3 (for example, consul-F2_1.20.4-1_amd64.deb becomes consul-F3_1.20.4-1_amd64.deb). Pipelines that pin FIPS artifact names or version strings must be updated.SECURITY:
Two mitigations are applied:
Each RPC request header is now validated against RPCMaxHeaderBytes (default 512 bytes) before it is decoded. Every length prefix in the header is checked against the limit, so an oversized value is rejected before the decoder allocates memory for it, and the connection is closed before any ACL evaluation. Request bodies remain unbounded by this limit.
A per-request read deadline (reusing RPCHandshakeTimeout) is applied inside handleConsulConn and handleInsecureConn so that a slow attacker trickling an oversized header cannot retain a goroutine and logical heap indefinitely.
service:write or node:write could delete peer-imported catalog
objects by supplying a non-default PeerName in a Catalog.Deregister request.
Authorization was checked only against the local service or node name, not the peer
origin, allowing deletion of objects in a peer-scoped catalog namespace the caller
does not control. Catalog.Deregister now rejects any request whose PeerName is
not the default, mirroring the existing guard on Catalog.Register and
Catalog.ListServices. Legitimate peer-state deletion continues through the internal
PeeringBackend.CatalogDeregister path.IMPROVEMENTS:
GOFIPS140=v1.0.0, CMVP Certificate #5247). The runtime FIPS line now reports FIPS 140-3 Enabled, crypto module v1.0.0. FIPS builds no longer require cgo or a vendored Go toolchain. FIPS 140-2 and FIPS 140-3 agents are permitted to join the same cluster; rolling upgrades from +ent.fips1402 to +ent.fips1403 are supported.BUG FIXES:
BREAKING CHANGES:
+ent.fips1402 to +ent.fips1403, and package and container artifacts change from the F2 suffix to F3 (for example, consul-F2_1.20.4-1_amd64.deb becomes consul-F3_1.20.4-1_amd64.deb). Pipelines that pin FIPS artifact names or version strings must be updated.SECURITY:
Two mitigations are applied:
Each RPC request header is now validated against RPCMaxHeaderBytes (default 512 bytes) before it is decoded. Every length prefix in the header is checked against the limit, so an oversized value is rejected before the decoder allocates memory for it, and the connection is closed before any ACL evaluation. Request bodies remain unbounded by this limit.
A per-request read deadline (reusing RPCHandshakeTimeout) is applied inside handleConsulConn and handleInsecureConn so that a slow attacker trickling an oversized header cannot retain a goroutine and logical heap indefinitely.
service:write or node:write could delete peer-imported catalog
objects by supplying a non-default PeerName in a Catalog.Deregister request.
Authorization was checked only against the local service or node name, not the peer
origin, allowing deletion of objects in a peer-scoped catalog namespace the caller
does not control. Catalog.Deregister now rejects any request whose PeerName is
not the default, mirroring the existing guard on Catalog.Register and
Catalog.ListServices. Legitimate peer-state deletion continues through the internal
PeeringBackend.CatalogDeregister path.IMPROVEMENTS:
GOFIPS140=v1.0.0, CMVP Certificate #5247). The runtime FIPS line now reports FIPS 140-3 Enabled, crypto module v1.0.0. FIPS builds no longer require cgo or a vendored Go toolchain. FIPS 140-2 and FIPS 140-3 agents are permitted to join the same cluster; rolling upgrades from +ent.fips1402 to +ent.fips1403 are supported.BUG FIXES:
BREAKING CHANGES:
+ent.fips1402 to +ent.fips1403, and package and container artifacts change from the F2 suffix to F3 (for example, consul-F2_1.20.4-1_amd64.deb becomes consul-F3_1.20.4-1_amd64.deb). Pipelines that pin FIPS artifact names or version strings must be updated.SECURITY:
Two mitigations are applied:
Each RPC request header is now validated against RPCMaxHeaderBytes (default 512 bytes) before it is decoded. Every length prefix in the header is checked against the limit, so an oversized value is rejected before the decoder allocates memory for it, and the connection is closed before any ACL evaluation. Request bodies remain unbounded by this limit.
A per-request read deadline (reusing RPCHandshakeTimeout) is applied inside handleConsulConn and handleInsecureConn so that a slow attacker trickling an oversized header cannot retain a goroutine and logical heap indefinitely.
service:write or node:write could delete peer-imported catalog
objects by supplying a non-default PeerName in a Catalog.Deregister request.
Authorization was checked only against the local service or node name, not the peer
origin, allowing deletion of objects in a peer-scoped catalog namespace the caller
does not control. Catalog.Deregister now rejects any request whose PeerName is
not the default, mirroring the existing guard on Catalog.Register and
Catalog.ListServices. Legitimate peer-state deletion continues through the internal
PeeringBackend.CatalogDeregister path.IMPROVEMENTS:
GOFIPS140=v1.0.0, CMVP Certificate #5247). The runtime FIPS line now reports FIPS 140-3 Enabled, crypto module v1.0.0. FIPS builds no longer require cgo or a vendored Go toolchain. FIPS 140-2 and FIPS 140-3 agents are permitted to join the same cluster; rolling upgrades from +ent.fips1402 to +ent.fips1403 are supported.BUG FIXES:
Port field in RPC and streaming subscription responses for multiport services. Older agents that cannot decode NodeService.Ports now correctly receive the default port value. The backfill logic is applied at RPC and subscription adapter boundaries via the shared multiport adapter so it is used consistently across Catalog, Health, PreparedQuery, and streaming subscription responses. [GH-23833]