Kong AI Gateway changelog

Uses: AI Gateway
Related Documentation
Tags

Changelog for supported Kong AI Gateway versions.

This is the changelog for the on-prem Kong AI Gateway runtime. policies are a control plane concept. In the runtime they’re implemented as plugins, which is the terminology you’ll see in this changelog.

2.3.2

Release date 2026/11/19

Bugfix

Core

  • Applied upstream nginx security patches for CVE-2026-123633.

  • Applied upstream nginx security patches for CVE-2026-123007.

2.3.1

Release date 2026/11/10

Bugfix

Core

  • Applied upstream nginx security patches for CVE-2026-42533.

  • Applied upstream nginx security patches for CVE-2026-48142.

  • Applied upstream nginx security patches for CVE-2026-56434.

  • Applied upstream nginx security patches for CVE-2026-60005.

  • Applied upstream nginx security patches for CVE-2026-9256.

  • Applied upstream nginx security patches for CVE-2026-42055.

2.3.0

Release date 2026/11/04

Breaking Change

Plugin

  • ai-nvidia-nemo-guardrail:

    • The default config.auth.prefix is now empty instead of Bearer , so configurations that relied on the default send the raw API key; set config.auth.prefix explicitly when a scheme is required. The response guard also evaluates the full conversation instead of only the assistant answer, which can block responses that were previously allowed.

  • graphql-proxy-cache-advanced:

    • Changed the cache_control default to true. The plugin now caches only responses with upstream freshness information. To cache by cache_ttl, set it to false.

Feature

Plugin

  • AI Gateway:

    • AI analytics and logs now report input_tokens/output_tokens and their _details breakdowns for every GenAI category, including text generation and embeddings. The legacy prompt_tokens/completion_tokens pair is unchanged and is still reported for text generation and embeddings, so existing consumers and dashboards keep working, but it is deprecated in favour of the canonical names.

      The analytics payload also carries token data the gateway already measured but previously dropped: prompt-cache writes (cache_write_tokens and its per-TTL breakdown cache_write_tokens_by_ttl), video token counts, image tokens in the output and cached breakdowns, and the provider-reported service_tier.

      AI logs and analytics now also split the request cost into input_token_cost and output_token_cost next to cost. The input side covers fresh, cache-read and cache-write tokens, both sides carry the context-window and service-tier multipliers, and the two add up to cost.

  • access-policy-engine:

    • Added a new plugin that enforces the Cedar policy set carried by the authenticated principal, compiled from the access-policy-engine Rust library (built as libape with LuaJIT FFI bindings). Requests without a Kong Identity principal pass through; enforced requests fail closed. Model calls are enforced as CallModel; MCP traffic is enforced only with enforce_mcp (off by default), tool calls as CallTool and every other MCP request as connect on the server; other routes pass. The control plane’s RequestContext envelope is the context record. WebSocket upgrades are enforced at the handshake. Key-auth and openid-connect now carry the principal’s group memberships and effective policy set on authenticated_principal. The Konnect analytics payload gains an ape entry with the decision (allow, deny, error) and the list of determining policies.

    • Aligned the Cedar request document with the groups contract change (pin 7a6dd65): memberships ride the identity’s top-level groups; type is an entity attribute.

  • ai-guardrails:

    • Added log_blocked_content field to log blocked payloads in ai-prompt-guard, ai-semantic-prompt-guard, and ai-semantic-response-guard.

  • ai-mcp-proxy:

    • Added allow_when and deny_when CEL expressions to default_acl entries and per-tool acl, so tool access can be decided from Kong Identity principal attributes, consumer groups, headers and other request context.

    • Added configurable credential placement for Token Vault providers. The new auth.token_vault.credential record configures where the gateway reads the caller’s inbound credential and where it writes the vault-issued outbound credential. Each placement accepts a header, a query parameter, or a body field. The gateway rejects the request when the request shape cannot carry the placement, such as with LLM auth integrations and MCP tool-call bodies. MCP discovery carries headers only, so a query or body placement is skipped there with a warning and discovery runs unauthenticated; the tools/list gate likewise ignores a body placement with a warning, and that provider’s tools stay hidden. The schema no longer accepts the server.token_vault field. Set auth.token_vault instead.

    • Added gateway-hosted OAuth enrollment for Token Vault providers: credentials are envelope-encrypted with a gateway-held key pair and only the encrypted blob is stored.

      Credential storage is configured through the new required auth.token_vault.storage record (set storage.current_version and storage.keys for the key versions; rotate by adding a version and flipping current_version); the L2 exchange cache settings moved to auth.token_vault.cache.redis, and the interim mode field was removed.

    • Added per-method and per-tool rate limiting via the new rate_limiting config field.

    • Migrated plugin to use shared /_vault/callback[/initiate] endpoints. token-vault-callback-handler is now a prerequisite for this plugin.

  • ai-proxy-advanced:

    • Added an http_5xx option to failover_criteria that matches any 5xx upstream status, including statuses nginx does not natively retry such as 501 or 522. Matched 5xx responses count toward the target circuit breaker failure threshold.

    • Added cost tag capture from ai-cost-* request headers and user-configured header names and prefixes in Konnect. Tags appear in analytics custom fields with the ai_cost. prefix.

    • Added dynamic pricing: a node-level pricing catalog in kong.conf that ai-proxy-advanced consults for per-request cost. dynamic_pricing (on by default) enables it for the node; dynamic_pricing_uri polls every 60 seconds (http(s):// or air-gapped file://); when unset, Konnect gateways fetch their org catalog with the cluster certificate and others use the public baseline. Pricing resolves per target: any model.options pricing field (rates, cost lists, factors) makes model.options price that target on its own; every other target is priced by the catalog, and an unpriced model bills 0 with a rate-limited warn. ai-semantic-cache prices cache savings from the same per-target source; ai-rate-limiting-advanced’s cost pre-charge does too, so a catalog-priced target reserves in-flight quota at its catalog rate. While enabled, the pricing partial is bypassed; with dynamic_pricing = off it resolves from the partial first, then model.options.

    • Added http_400, http_401, http_402, http_405, http_408, http_409, http_413, http_415, http_422, http_522, and http_529 to the supported failover_criteria values. Configured HTTP 429 and 5xx responses count toward the target circuit breaker failure threshold; other 4xx responses trigger failover without counting as failures.

    • Added native passthrough for the Gemini and Vertex Interactions API.

    • Added support for Baseten provider for chat completions, with Model APIs as the default upstream and dedicated deployments via upstream_url.

    • Added support for the Bedrock Runtime native OpenAI and Anthropic APIs at /openai/v1/chat/completions, /openai/v1/responses, and /anthropic/v1/messages.

    • Added support for Bedrock Mantle models, for four endpoints and llm_formats: /v1/chat/completions (openai), /v1/responses (openai), /v1/files (openai), and /v1/messages (anthropic).

    • Added vpc_endpoint Bedrock option to send requests to a custom VPC endpoint (host or host:port; the port defaults to 443).

    • Anthropic’s “preserved thinking” binds thinking blocks to the conversation prefix that produced them, so replaying a thinking block after the preceding context changed fails upstream with HTTP 400. Kong now enables the drop_block binding mode (with its required thinking-binding-controls-2026-08-01 beta flag) by default on every outgoing request that carries thinking, for the preserved-thinking generation and newer — Fable and Mythos in any release, and Sonnet, Haiku, and Opus from 5.5 — across Anthropic, Google Cloud Vertex AI, Azure Foundry, and Amazon Bedrock.

    • Dynamic pricing now bills server-side tool calls (web search, web fetch, file search and remote MCP) at the per-call prices in the catalog’s provider-level tools list. The tool cost is added to usage.cost outside every token pricing multiplier, and is broken down per tool in the new usage.tool_cost and usage.tool_usage log fields.

  • dynamic pricing:

    • Added catalog entry overrides: an entry with override: true and an effective_to RFC 3339 timestamp (YYYY-MM-DDTHH:MM:SS[.fff]Z, Kong AIP 142; fractional seconds truncated) bills its pricing until the gateway’s clock reaches effective_to, then falls back to the entry’s baseline_price (same rate shape as pricing). A malformed combination (unparseable effective_to, missing or invalid baseline_price, or either field without override: true) degrades to a plain override entry with a warn instead of dropping it, so the request never bills 0. effective_from stays catalog-side: the catalog publishes one entry per (provider, model) with last-write-wins, so the data plane never picks between entries by time.

  • token-vault-callback-handler:

    • Added a new plugin to handle all Token Vault OAuth callbacks at the global well-known routes /_vault/callback[/initiate].

  • token-vault-credential-manager:

    • Added a new plugin that provides a self-serve connectors UI backed by Kong Identity’s Token Vault, allowing you to list, connect, and revoke third-party provider credentials.

  • token-vault-exchange:

    • Added a new plugin that exchanges your bearer token for an upstream credential via the Kong Token Vault, configurable on arbitrary routes.

    • Migrated enrollment-required 401s to surface the gateway-hosted /_vault/callback/initiate URL; token-vault-callback-handler is now a prerequisite for this plugin.

  • ai-prompt-compressor:

    • Added ai.compressor.request_bytes_before and ai.compressor.request_bytes_after to the log for native request bodies. They show savings on system and tools that Headroom’s token counts don’t include.

    • Added prompt compression metrics to Konnect Analytics and the response headers with the result.

    • Added x-opencode-session-id to the default session_id_headers list. opencode v2 sends X-Session-Id with the parent session id for subagents. With the new default, a subagent and its parent no longer share one Headroom session.

  • ai-rate-limiting-advanced:

    • Added the principal policy match type, which matches or partitions on the Kong Identity principal’s id, display_name or a metadata.<name> scalar.

    • Added the expressible custom_key field. It replaces the subject part of every matched policy’s counter key for the request and keeps model and provider partitions. Set expressions.custom_key to compute it per request with CEL.

  • graphql-proxy-cache-advanced:

    • Added response_code, content_type, and cache_control configuration options to control which upstream responses the plugin caches.

  • openid-connect:

    • Added actor token support to token exchange. Configure token_exchange.request.actor_token to include an actor token, read from a request header or a static token value, in the token exchange request.

    • Added credential_claims to support mapping virtual credentials from multiple claim paths, while retaining and supporting the existing credential_claim as a shorthand configuration.

    • Added principals.jit.all_claims and principals.jit.claims to share token claims with the Kong Identity principal lookup. all_claims sends every verified token claim. claims sends only the named claims. Kong Identity maps the claims through the trusted IdP’s JIT mapping. This mapping provisions or updates the principal just in time. Kong Identity owns that setting. By default the lookup carries only the principal claim (e.g. sub).

    • Added protected_resource_metadata configuration to support RFC 9728 OAuth 2.0 Protected Resource Metadata, enabling automatic authorization server discovery by serving a well-known metadata document and enriching WWW-Authenticate challenge headers with resource_metadata and scope attributes.

    • Added support for Microsoft Entra ID’s On-Behalf-Of flow through a new token_exchange.grant_type = jwt_bearer (RFC 7523) mode, plus map_identities_from and token_exchange.provider.

    • Added the token_exchange.subject_token_issuers[].deny_on_mismatch config option. When set to true, Kong denies a bearer token whose issuer matches but whose conditions do not match, instead of passing it through unexchanged. Kong also logs this when it happens.

    • Added verify_signature and jwks_uri fields to token_exchange.subject_token_issuers for verifying incoming subject token signatures before exchange.

  • opentelemetry:

    • Added a metrics.enable_credential_attribute config option. When enabled, the kong.auth.credential.id attribute is added to the http.server.request.count, http.server.request.size, http.server.response.size, gen_ai.client.token.usage and kong.gen_ai.llm.cost metrics when an authenticated credential is available. The new opentelemetry_metrics_credential_attr_limit configuration property sets the maximum number of distinct attribute values for each worker.

    • Added a metrics.enable_principal_attribute config option. When enabled, the principal attributes are added to the metrics when an authenticated principal is available.

  • proxy-cache-advanced:

    • Added cache_by_principal configuration option to switch cache key segmentation by Principal when using Kong Identity Directory. When enabled, the authenticated Principal’s id is used to compose the cache key.

  • proxy-cache:

    • Added cache_by_principal configuration option to switch cache key segmentation by Principal when using Kong Identity Directory. When enabled, the authenticated Principal’s id is used to compose the cache key.

  • rate-limiting-advanced:

    • Added expressions.limit to drive individual limit entries via CEL, falling back to the static value on error.

    • Added support for rate limiting by Kong Identity Directory Principal, via identifier: principal, and for overriding the rate-limiting key entirely via the new, expressible custom_key field.

    • Added the counter_key field to let you choose which consumer attribute is now used as the rate-limiting counter identifier. Supported values are consumer.id, consumer.username, and consumer.custom_id.

    • added the missing_identifier_behavior config option. It controls what the plugin does when identifier cannot resolve a value for a request. Set it to ip (default) to fall back to the client IP address, reject to respond with 403 Forbidden, skip to let the request through uncounted, or fixed_key to count the request under one shared, fixed key.

  • http-log, file-log, tcp-log, udp-log, syslog, loggly, solace-log, kafka-log: Added an opt-in log_sanitization configuration block. When enabled, sensitive data are redacted in log entry.

Core

  • Added support for driving certain plugin configuration fields with a CEL expression, evaluated per request and falling back to the field’s static value if unset or if evaluation fails.

  • analytics: Added the groups of the authenticated Kong Identity principal (id and type) to the analytics payload as principal.groups.

  • Bundled the KFC Rust shared library for use by Lua FFI integrations.

  • Konnect Only. Debugger: Increased max body size from 128KB to 1MB for debugger payload capture feature.

  • Plugin conditions (and CEL-based ACL rules) can now reference principal.groups and principal.type. The first is the list of groups for the authenticated principal. The second is the principal kind, group or principal.

Configuration

  • Added the ai_pricing_cache_size property to set the size of the shared memory zone that stores the AI pricing catalogs polled from dynamic_pricing_uri.

PDK

  • Added the principal field to the log serialization output. The field reports the id and display_name of the authenticated principal for the request.

  • jwe: Added support for the RSA-OAEP-256 key management algorithm in JWE.

Bugfix

Plugin

  • AI Guardrails:

    • Fixed an issue where the prompt text the guardrail plugins match against ended on a blank line, so a deny or allow pattern anchored to its end (\z, \Z or $) never matched. The \n\n separator now goes only between the included messages, and a message whose multi-modal parts carry no text no longer puts one in front of the message that follows it. A pattern written against that blank line instead, one ending in \n\n, \n+$ or \s+$, stops matching once this ships: a deny pattern of that shape forwards content it used to block, and an allow pattern of that shape rejects content it used to pass. Re-anchor it on the content (\z or $), or keep its trailing whitespace optional (\s*$). Plugins that read this text build a slightly different string than before: ai-semantic-cache entries stored before the upgrade can miss once, and the ai-rate-limiting-advanced pre-charge is one token lower per request.

  • ai-a2a-proxy:

    • Fixed an issue where a JSON-RPC endpoint whose path ended with /extendedAgentCard had its advertised endpoint URL truncated.

    • Fixed an issue where empty arrays in agent cards were re-encoded as empty objects, which failed strict A2A SDK schema validation.

  • ai-guardrails:

    • Fixed the continue_on_detection description on ai-prompt-guard, ai-semantic-prompt-guard and ai-semantic-response-guard to match their behavior.

  • ai-mcp-proxy:

    • Fixed a bug where input_cost or output_cost set to 0, would cause all LLM costs to output as 0.

    • Fixed an issue where failed upstream tools/list requests caused immediate retries and listener errors under load.

    • Fixed an issue where token_vault upstream auth answered 500 instead of 502 when the token vault couldn’t supply a credential.

    • Fixed an issue where arrays of RPC messages were not blocked, potentially leading to ACL check bypass.

    • Fixed an issue where modern MCP protocol version metadata was forwarded to legacy upstream servers.

    • Fixed an issue where a client could pass Kong-internal headers such as X-Kong-MCP-Ctx to the internal tool-call request and redirect it in listener mode.

    • Fixed an issue where path traversal could be done with embedded dot segments in tool path parameters.

    • Fixed an issue where notifications/tools/list_changed was not delivered on the call that completes Token Vault enrollment, so clients that accept SSE responses can now refetch their tool list without holding a subscription stream.

    • Fixed an issue where upstream-server mode skipped MCP servers on routes with only regex paths. Internal calls are now pinned to that route, not re-matched by path.

    • Fixed an issue where Kong kept replaying an upstream MCP session the server had already dropped, serving an empty or stale tool list instead of starting a new session.

    • Fixed an issue where token_vault incorrectly returned a 500 error for missing or rejected bearer tokens. Now returns 401 with WWW-Authenticate header.

    • Improved the Token Vault enrollment prompts served to LLM callers: the virtual token_vault_authenticate/token_vault_check_authentication_status tools now state upfront they are the only tools available until authorization completes, and the enrollment URL is delivered with strict verbatim-copy instructions so agents relay it to the user’s browser unmodified.

    • tools[].description is now optional in passthrough-listener and upstream-server modes; it remains required in conversion-listener and conversion-only modes.

  • ai-model-selector:

    • Fixed an issue where a model resolved by ai-model-selector plugin from a custom location specified was never written into the upstream request body when the provider is identical with the adapter, causing the provider to reject the request for a missing model field.

  • ai-proxy and ai-proxy-advanced:

    • Fixed an issue where reading back a stored object (for example GET /v1/responses/{id}) counted its token usage and server-side tool calls again, billing the response a second time in usage.cost and in cost-based rate limiting. A non-POST request to a RESTful object API (responses, assistants, files, batches) is no longer metered, except the poll that returns the result of a background Responses job, whose POST carries no usage; the batch output file download and realtime sessions are still metered.

  • ai-proxy, ai-proxy-advanced, ai-a2a-proxy, ai-mcp-proxy:

    • Fixed an issue where requests sent through proxy_config failed with a 503 when the data plane could not resolve the upstream host itself; Kong now leaves upstream name resolution to the forward proxy.

  • ai-proxy, ai-proxy-advanced:

    • Fixed an issue where downloading a non-JSONL file through the llm/v1/files content endpoint with logging.log_statistics enabled truncated the response body, because the batch usage parser aborted the body filter partway through the download.

    • Fixed an issue where an Anthropic-format request sent to an Anthropic-format target (Anthropic, or Bedrock with a Claude model) lost the client’s cache_control TTL when a plugin modified the request body. Kong replaced the TTL with a model default of 1h, which bills cache writes at the one-hour rate. Kong now keeps each client cache_control that the rewritten body still carries. A breakpoint that survives without the client’s value gets Anthropic’s five-minute default, and a breakpoint the plugin removed is not recreated. On the rebuilt body, a ttl the provider rejects now reaches the provider unchanged and comes back as the provider’s error, instead of being replaced by a Kong default.

    • Fixed the OpenAI-shaped SSE chunks that Kong synthesizes from a Gemini stream breaking the OpenAI streaming contract, so stock OpenAI SDKs could not read them. The chunks carried no object discriminator, reported created as 0 because Gemini omits createTime on stream chunks, repeated role in every delta instead of the first one only, and merged finish_reason into the last content-bearing chunk instead of sending the closing chunk with an empty delta that OpenAI clients expect. Passthrough streams were never affected.

  • ai-proxy:

    • Fixed an issue where a Bedrock target without an auth block accepted a client Authorization: Bearer bedrock-api-key- header, so any caller could replace the gateway credentials with its own key. The override now requires the target to set auth.allow_override to true.

    • Fixed an issue where a Bedrock target refused a client Authorization: Bearer bedrock-api-key- header with HTTP 500 instead of a 4xx, which reported a caller-caused rejection as a server fault. The refusal is now HTTP 403.

    • Fixed an issue where a client x-api-key header was forwarded to AWS on a Bedrock target.

    • Fixed an issue where ‘OpenInference’ OpenTelemetry keys format would output token counts as float/double type, resulting in 0 token counts, depending on the backend collector.

    • Fixed an issue where it was not possible to retrieve Anthropic batch inference results when presented in JSONL content type.

    • Fixed an issue where the Bedrock Authorization: Bearer bedrock-api-key- override header was matched case-sensitively, regarding the allow_override auth configuration.

  • ai-proxy-advanced:

    • Fixed an issue where Gemini reasoning content was not emitted as thinking blocks in Anthropic-format responses. Requests can now opt out with thinking.display.

    • Fixed an issue where request bodies were re-encoded with a non-deterministic JSON key order; content-neutral transforms now forward the client’s original request bytes.

    • Fixed an issue where upstream SSE event: lines were dropped from streaming responses for the llm/v1/responses route type.

    • Fixed an issue where driver response headers (e.g. Set-Cookie) were not removed from streaming responses.

    • Fixed an issue where targets kept using expired credentials after a vault rotated the secrets referenced in their auth until the configuration was reloaded.

    • Fixed an issue where a forced tool_choice was rejected by Claude Sonnet 5.5. The plugin now sends auto to a model that refuses a forced tool call.

    • Fixed an issue where an Anthropic-native request to Vertex carried temperature, top_p, top_k or seed for a model that rejects them. The plugin now drops them on that route, as it already did for OpenAI-format requests.

    • Fixed an issue where an Anthropic target rejected OpenAI-format history that replayed the reasoning field.

    • Fixed an issue where anthropic-format tool call conversion dropped parallel tool calls, emitted non-object tool_use input, and dropped text alongside tool results.

    • Fixed an issue where replaying history to a Cohere target was not restored.

    • Fixed an issue where the gen_ai.system_instructions OpenTelemetry span attribute contained the system prompt even when logging.log_payloads was disabled.

    • Fixed an issue where Databricks rejected chat history that replayed the reasoning field returned by Kong.

    • Fixed an issue where DeepSeek in thinking mode rejected tool-call history without reasoning content.

    • Fixed an issue where empty, null, or malformed tool call arguments produced a missing or invalid tool input in Anthropic and Bedrock requests instead of {}.

    • Fixed an issue where Gemini 3 models rejected OpenAI-format requests with tool-call history because the converted function calls had no thought signature.

    • Fixed an issue where a Mistral target rejected OpenAI-format history that replayed the reasoning field.

    • Fixed an issue where Bedrock streaming through invoke-with-response-stream reported no cache read/write token counts on bedrock-format responses and no per-TTL cache-write breakdown, which billed cached tokens at the full input rate.

    • Fixed an issue where Databricks reasoning was dropped when the upstream returned it as reasoning_content instead of content parts.

    • Fixed an issue where the Anthropic adapter dropped image content blocks and leaked non-text tool_result blocks when converting requests to the OpenAI format.

    • Fixed an issue where empty tool_use input was converted to [] in Anthropic adapter.

    • Fixed an issue where unsigned thinking blocks from open-weight Anthropic-compatible models were stripped from replayed request history, causing models to re-reason every turn.

    • Fixed an issue where using Anthropic models over Bedrock would produce redundant error-level spam in the Kong logs.

    • Fixed an issue where reasoning requests to Amazon Nova models with a high reasoning effort failed with a 400 error because Bedrock rejected the combination of reasoningConfig.maxReasoningEffort = "high" with inferenceConfig.maxTokens, temperature, or topP. Those sampling parameters are now dropped from the request for such reasoning requests.

    • Fixed an issue where a request body the gateway rebuilt, or that a filter rewrote on a route without a driver transform, was re-encoded in a member order that changed between workers and restarts, causing upstream prompt cache misses.

    • Fixed an issue where a gateway-derived reasoning_effort outside the levels an OpenAI-family target model accepts was forwarded as it was, which made the provider answer 400. A derived value above the model’s ceiling is now lowered, and one below its floor is raised.

    • Fixed an issue where temperature, top_p, top_k and seed were sent to Claude models that reject them, so your requests to those models no longer fail.

    • Fixed response_format JSON schema requests failing on Claude 4.5 and later models. For an OpenAI-format request the plugin strips the JSON Schema keywords the Anthropic API rejects, and answers 400 for a schema it cannot express.

  • token-vault-credential-manager:

    • Reduced Kong Identity / Token Vault load from the connectors page: the vault provider list is now cached per node for 30s (failed fetches negatively cached for 2s) instead of being re-fetched on every poll, and the page now backs off when per-provider probes fail instead of polling at full rate.

    • The connectors endpoint now probes vault providers concurrently instead of sequentially, so its latency is the slowest single probe rather than the sum (previously K providers x 5s timeout against a hung vault), and the page’s base poll interval is now 30s with an immediate poll when the tab regains focus (e.g. returning from the enrollment flow).

  • ai-aws-guardrails:

    • Fixed an issue where streaming responses could end before pending asynchronous guardrail checks completed, losing their moderation results from request logs.

    • Fixed an issue where masking dropped the request payload on routes with a non-OpenAI llm_format.

  • ai-gcp-model-armor:

    • Fixed an issue where the OUTPUT guardrail failed to detect prompt injection in Gemini responses by extracting the answer text before sending it to Model Armor.

  • ai-prompt-compressor:

    • Fixed a 500 error when a Headroom call reused a pooled connection that Headroom had already closed. Kong now sends the call once more on a new connection.

    • Fixed an issue where compression stayed off on a route where the client format and the provider format match through a driver shortcut (an Anthropic body to Anthropic models on Bedrock). The compressor read the route type from its own configuration, which has no such field. It also no longer replaces the body the provider driver builds, so driver defaults (for example max_tokens) survive on the wire. A body whose conversation the round trip left unchanged is forwarded exactly as the client sent it, and a round trip that changed a tool result alone is written back.

    • Fixed an issue where a compressor backend that answered with an empty text list failed every request on the route with an HTTP 500. An empty answer means nothing was compressed, so the plugin now forwards the request to the upstream uncompressed, with no X-Kong-Prompt-Compression-* headers and no compression metrics.

    • On routes to Anthropic, Kong now forwards the anthropic-beta tokens that Headroom returns, after the client’s own. Headroom Enterprise’s tool search needs this header for the tools it defers.

    • The Headroom provider now skips native Bedrock Converse requests and logs a warning. Headroom can’t read Converse content blocks, so these requests came back uncompressed with an extra model field.

  • ai-prompt-template:

    • Fixed an issue where a templated request on an ai-proxy or ai-proxy-advanced route failed instead of being rendered. And the guardrail plugins run after ai-prompt-template and ai-prompt-decorator, so a guard inspects the prompt the provider receives.

  • ai-rate-limiting-advanced:

    • Fixed an issue where calendar window policies named the window segment of rate-limit response headers nil instead of the configured period (week, month). Header handling remained incomplete when an ai-proxy-advanced balancer denies one target and then serves a later target: rate-limit headers stored for the denied target can still remain on the successful response.

  • ai-sanitizer:

    • Fixed an issue where credentials was dropped from the anonymize list when all was also selected, which silently turned off credential sanitization.

  • ai-semantic-cache:

    • Fixed an issue where, with ai-sanitizer and recover_redacted enabled on the same route, a response with one caller’s restored PII could be cached and served to other callers. Such responses are no longer cached.

    • Fixed an issue where the plugin cached a request with an Authorization header when cache_control was enabled. The plugin read a wrong nginx variable, so it never found the header.

    • Fixed an issue where a streaming cache hit for the llm/v1/responses route type replayed SSE frames without event: lines.

  • ai-semantic-response-guard:

    • Fixed an issue where the OUTPUT guard failed to block responses from non-OpenAI providers (e.g. Bedrock) because it embedded the raw provider response envelope instead of the answer text. The guard now extracts the answer text with the shared driver extractors, for all supported providers.

  • graphql-proxy-cache-advanced:

    • Fixed an issue where responses to authenticated requests could be incorrectly cached and served to other callers.

    • Fixed an issue where the plugin’s cache key could collide, causing cache entries to be shared across different requests.

    • Fixed an issue where the plugin’s cache key did not include the request’s Host header, causing responses to be shared across different hosts on the same route.

    • Fixed an issue where the plugin’s cache key did not include the authenticated consumer, so cache entries could be shared across different consumers.

    • Fixed an issue where the plugin cached responses that should not have been cached.

  • graphql-rate-limiting-advanced:

    • Fixed an issue where pass_all_downstream_headers forwarded stale Content-Length, Content-Encoding, and Transfer-Encoding headers, breaking schema introspection.

  • jwe-decrypt:

    • Fixed an issue where an invalid alg or enc value in the JWE header caused a 500 error. It now returns a 403 error, and a missing alg value now returns a 400 error instead.

  • oauth2-introspection:

    • Fixed an issue where the credential ID could expose sensitive information.

  • openid-connect:

    • Fixed an issue where plugin with no_proxy configured still sent a malformed absolute-form request to the IdP.

    • Fixed an issue where claim headers added to upstream requests remained unavailable to later plugins after an earlier read returned no value.

    • Fixed an issue where token_exchange.cache defaulted to disabled instead of its documented enabled default when omitted.

    • Fixed an issue where a session was not saved on a request that performed a token exchange. A session-restored request now reuses the exchanged identity instead of re-exchanging.

    • Fixed an issue where token_exchange scope and audience conditions matched only the claims scope and aud. Support for scopes_claim/audience_claim is now added to token_exchange.

    • Fixed an issue where an invalid JSON response from the JWKS discovery endpoint caused an unhandled error instead of a proper error message.

    • Fixed proof_of_possession_mtls: optional to verify only tokens with the certificate hash claim. Cert-bound tokens now require a valid cert, others are not mTLS-verified.

    • Fixed the token exchange cache to stop outliving its subject token, mixing entries across grant_type/provider/client, and serving stale results after a plugin update under incremental config sync. Use cache_tokens_salt to scope the cache further.

  • opentelemetry:

    • Fixed an issue where multi-valued HTTP headers in access logs were exported as numbered attributes instead of an array, which caused Elasticsearch mapping conflicts and dropped records.

    • Fixed an issue where the kong.shared_dict.size metric reused the same datapoint timestamp across exports, causing monitoring systems to treat the series as stale.

    • Fixed an issue where the plugin removed existing metric series when the otel_metrics shared dictionary was full. The plugin now drops the new series and keeps the existing ones.

    • Fixed an issue where unchanged metrics kept stale timestamps, which caused Prometheus remote write requests to reject metrics batches.

  • proxy-cache-advanced:

    • Fixed an issue where responses to authenticated requests could be incorrectly cached and served to a caller without valid credentials.

    • Fixed an issue where the plugin could generate the same cache key for different requests, causing Kong to serve a cached response meant for a different request.

    • Fixed an issue where the plugin’s cache key did not include the request’s hostname, causing Kong to serve a cached response meant for a different hostname.

    • Fixed an issue where the cache key did not correctly segment by consumer group.

  • proxy-cache:

    • Fixed an issue where responses to authenticated requests could be incorrectly cached and served to other callers.

    • Fixed an issue where the plugin generated the same cache key for different requests. Kong then served a cached response meant for a different request.

    • Fixed an issue where the plugin’s cache key did not include the request’s hostname, causing Kong to serve a cached response meant for a different hostname.

    • Fixed an issue where the cache key did not correctly segment by consumer group.

  • request-callout:

    • Fixed an issue where the plugin’s cache key did not vary by all of the callout request’s headers, causing Kong to serve the wrong cached response.

    • Fixed an issue where the plugin could generate the same cache key for different requests, causing Kong to serve a cached response meant for a different request.

  • saml:

    • Fixed an issue where the plugin did not enforce the SAML assertion’s audience restriction. The plugin now rejects assertions whose audience does not match the configured issuer. Note: the stricter validation can break existing logins if Audience value is not an exact match for issuer.

  • vault-auth:

    • Fixed an issue where credentials remained usable after being rotated, revoked, or reassigned to another consumer on Vault KV v2 mounts.

  • ai-sanitizer, ai-semantic-response-guard: Fixed an issue where response text extraction did not recognize Cohere v1 chat responses, which carry the answer as a top-level “text” field. The whole response body was treated as the text to sanitize or guard. A dedicated extractor now reads the answer text from Cohere v1 chat responses.

Clustering

  • Allow AI Gateway control planes and data planes with matching major and minor versions to run different patch releases.

  • Fixed an issue where a config backup exported by a data plane could not be re-imported if it contained non-workspaceable entities such as ca_certificates or oic_issuers. The export incorrectly included a ws_id field for those entities, which the declarative config schema rejects as an unknown field.

  • Fixed an issue where data planes retried unrecoverable TLS certificate setup failures instead of loading their backup configuration.

  • Fixed an issue where hybrid data planes stopped configuration retries after a certificate verification failure during control plane rotation. Retries now use cycling backoff; Konnect data planes keep stopping retries.

  • Fixed an issue where incremental DP (sync v2) would not export config to fallback storage when the control plane uses periodic polling instead of notify_new_version.

  • Fixed an issue where the control plane could fail to update a data plane’s status if that data plane reconnected or disconnected while the update was in progress.

Core

  • Fixed an issue where a failed balancer creation leaked the per-upstream create lock, so every later creation for the same upstream stalled for 30 seconds.

  • Fixed an issue where a healthcheck counter value of 255 crashed health checker creation. Healthcheck counters (successes, tcp_failures, timeouts, http_failures) are now limited to 254.

  • Fixed an issue where a Postgres connection whose setup statements timed out was returned to the pool, causing subsequent queries on that connection to fail with an internal server error.

  • Fixed an issue where AI proxy plugins rejected a price of 0 in model.options.input_cost and in the input_per_token and output_per_token fields of config.pricing.models. The same price was accepted in other pricing fields. A price of 0 is now valid in these fields, and a target with both costs set to 0 is billed as 0 with no missing-price warning.

  • Fixed an issue where failed kong.balancer tracing spans did not include the failure reason.

  • Fixed an issue where plugin server processes were left running after the worker managing them was killed, allowing unused processes to accumulate and increase memory usage.

  • Fixed an issue where the Distroless images did not ship the IANA timezone database (tzdata), causing ai-rate-limiting-advanced calendar windows to reject every timezone (including UTC) as not a valid IANA timezone and Data Planes to fail CP-to-DP config sync.

  • Fixed an issue where the /status/ready API responded with HTTP 500 or 200 instead of 503 when some workers failed in initialization.

  • Fixed an issue where Token Vault refresh did not write refreshed admin-enrolled shared OAuth credentials back to Kong Identity for ai-mcp-proxy and token-vault-exchange.

  • Fixed an SQL injection in the LLM usage counters. A model name that came from a client request was written into the license_llm_data table without escaping. The counters now escape every value before they run the statement.

  • Fixed the request body being written before the filters that can change it, which broke Bedrock and SageMaker requests signed with AWS Signature Version 4.

  • Fixed the X-Forwarded-Path header to hold a normalized and percent-escaped path. The gateway now removes dot segments and escapes unsafe bytes, such as CR and LF, before it sends the header to the upstream service.

  • RBAC: Fixed an issue where a temporary database outage during an RBAC user’s group lookup cached a wrong value, breaking every later Admin API request from that user.

PDK

  • Fixed an issue where dynamic upstream TLS settings set via the PDK were lost during buffered proxying.

  • Fixed an issue where kong.service.request.set_body() doesn’t process boundary with = character.

  • Fixed an issue where the kong.request.get_body() function incorrectly parsed Content-Disposition name parameter if it is not double quoted.

Admin API

  • Fixed an issue where RBAC self-protection checks were not applied consistently in the Admin API.

  • GET /admins/:admin?generate_register_url=true now requires permission to create admins and full ownership of the target’s roles before returning a registration token and URL. Previously any caller with read access on /admins* could read the registration JWT of an invited admin and use it to set the account password on the unauthenticated /admins/register endpoint.

  • PATCH /admins/:admin no longer accepts the status field, and identity fields (username, custom_id, email, rbac_token_enabled) of a more-privileged admin can no longer be changed by a delegated caller. DELETE /admins/:admin and PATCH /admins/:admin/workspaces/:workspace are subject to the same check. Previously a delegated writer on /admins* could re-open an existing admin’s registration flow and take over the account.

  • Registering an invited admin via POST /admins/register now correctly sets the admin status to APPROVED. Previously the status update used reversed DAO arguments and silently left the admin in INVITED state.

  • The /admins/:admin/roles and /rbac/users/:user/roles endpoints now verify that the caller holds every role they assign to or remove from another principal (callers with full access may delegate any role). A delegated writer on /admins* can no longer grant roles such as super-admin to another admin, which previously allowed a low-privileged identity admin to escalate to full Admin API control.

Dependency

Core

  • Added Ubuntu 26.04 (Resolute Raccoon) and RHEL 10 FIPS images.

  • Bumped brotli to 1.2.1.

  • Bumped libexpat to 2.9.0.

  • Bumped lua-resty-cel from 0.4.0 to 0.5.0.

  • Bumped multipart from 0.5.11 to 0.5.12.

  • Bumped OpenSSL to 3.5.9.

  • Bumped PCRE2 to 10.49.

  • Bumped snappy to 1.3.1.

  • Kong AI Gateway Docker images are now built on Ubuntu 26.04 and RHEL 10, replacing Ubuntu 24.04 and RHEL 9.

  • Removed the libpcre3 package dependency from Kong AI Gateway .deb packages; Kong bundles PCRE2.

  • Removed the pcre package dependency from Kong AI Gateway .rpm packages; Kong bundles PCRE2 and RHEL 10 no longer ships classic PCRE.

Default

  • Updated Bazel to 8.7.0 and migrated the build to Bazel Modules (Bzlmod).

Performance

Core

  • Improved the performance of proxied requests by skipping the regular expression engine when URI path escaping finds nothing to escape.

2.2.2

Release date 2026/11/19

Bugfix

Core

  • Applied upstream nginx security patches for CVE-2026-123633.

  • Applied upstream nginx security patches for CVE-2026-123007.

2.2.1

Release date 2026/11/10

Feature

Plugin

  • proxy-cache-advanced:

    • Added cache_by_principal configuration option to switch cache key segmentation by Principal when using Kong Identity Directory. When enabled, the authenticated Principal’s id is used to compose the cache key.

  • proxy-cache:

    • Added cache_by_principal configuration option to switch cache key segmentation by Principal when using Kong Identity Directory. When enabled, the authenticated Principal’s id is used to compose the cache key.

Bugfix

Plugin

  • ai-semantic-cache:

    • Security: Fixed CVE-2026-92443 by not caching authenticated responses, preventing them from being served to other callers.

  • graphql-proxy-cache-advanced:

    • Security: Fixed CVE-2026-92558 by digesting every cache key field and component to a fixed-length value before joining them, preventing two different requests from resolving to the same cache key, and one caller’s cached response from being served to another, when header values contain the separator characters.

    • Security: Fixed CVE-2026-92545 by always including the request Host in the cache key, preventing a response generated for one Host from being served to another Host on routes that match more than one Host or have no Host constraint.

    • Security: Fixed CVE-2026-92443 by not caching authenticated responses, preventing them from being served to other callers.

    • Security: Partially fixed CVE-2026-92442 by including the authenticated consumer’s id in the cache key, preventing one caller’s GraphQL response from being served to another caller on routes reachable by more than one caller. Responses marked Cache-Control: private or no-store can still be cached; this is not addressed in this release.

  • proxy-cache-advanced:

    • Fixed an issue where the cache key did not correctly segment by consumer group.

    • Security: Fixed CVE-2026-92558 by digesting every cache key field and component to a fixed-length value before joining them, preventing two different requests from resolving to the same cache key, and one caller’s cached response from being served to another, when request values contain the separator characters.

    • Security: Fixed CVE-2026-92545 by always including the request Host in the cache key, preventing a response generated for one Host from being served to another Host on routes that match more than one Host or have no Host constraint.

    • Security: Fixed CVE-2026-92443 by not caching authenticated responses, preventing them from being served to unauthenticated callers.

  • proxy-cache:

    • Fixed an issue where the cache key did not correctly segment by consumer group.

    • Security: Fixed CVE-2026-92558 by fixing cache key generation, preventing cached responses from being served to the wrong request.

    • Security: Fixed CVE-2026-92545 by including the hostname in the cache key, preventing cached responses from being served across hostnames.

    • Security: Fixed CVE-2026-92443 by not caching authenticated responses, preventing them from being served to other callers.

  • request-callout:

    • Security: Fixed CVE-2026-123041 by including all callout request headers in the cache key, preventing cached callout responses from being served to requests with different headers.

    • Security: Fixed CVE-2026-92558 by fixing cache key generation, preventing cached responses from being served to the wrong request.

  • saml:

    • Security: Fixed CVE-2026-92439 by evaluating the audience XPath relative to the Conditions node, so the plugin now rejects assertions whose audience does not match the configured issuer, preventing a validly signed assertion issued for a different Service Provider from being accepted when the IdP shares its signing key across Service Providers.

      Note: The stricter validation can break existing logins if the Audience value is not an exact match for issuer.

Core

  • Applied upstream nginx security patches for CVE-2026-42533.

  • Applied upstream nginx security patches for CVE-2026-48142.

  • Applied upstream nginx security patches for CVE-2026-56434.

  • Applied upstream nginx security patches for CVE-2026-60005.

  • Applied upstream nginx security patches for CVE-2026-9256.

  • Applied upstream nginx security patches for CVE-2026-42055.

  • Security: Fixed CVE-2026-92441 in the gateway by normalizing and percent-escaping the X-Forwarded-Path header (removing dot segments) and percent-escaping X-Forwarded-Prefix, including bytes such as CR and LF, preventing header injection and request smuggling against upstream services when these headers are sent upstream.

  • Security: Fixed CVE-2026-92544 in the LLM usage counters by escaping every value before running the SQL statement, preventing SQL injection when a client-supplied model name is written to the license_llm_data table.

Admin API

  • Security: Fixed CVE-2026-105986 in the Admin API by applying RBAC self-protection checks consistently, preventing those protections from being bypassed.

Dependency

Core

  • Bumped libexpat to 2.9.0.

  • Bumped OpenSSL to 3.5.9.

  • Bumped PCRE2 to 10.49.

  • Bumped snappy to 1.3.1.

  • Security: Fixed CVE-2026-111497 in the bundled Brotli compression library by bumping brotli from 1.1.0 to 1.2.1, preventing an out-of-bounds write that could allow remote code execution when Kong compresses or decompresses data with Brotli.

Performance

Core

  • Improved the performance of proxied requests by skipping the regular expression engine when URI path escaping finds nothing to escape.

2.2.0

Release date 2026/09/30

Breaking Change

Core

  • AI Gateway: The FIPS artifact serves an explicit TLS 1.2 cipher list: four ECDHE suites with AES-GCM. It replaces the OpenSSL FIPS alias, which resolves to 28 suites and also admits static RSA key transport and SSLv3-era suites. A deployment moving from the retired kong-ai-gateway-fips package that relies on one of the dropped suites must set ssl_cipher_suite = custom and ssl_ciphers. TLS 1.3 ciphersuites are unaffected; the global OpenSSL FIPS configuration sets those.

  • AI Gateway: The FIPS artifacts are renamed to state their certification level. The kong-ai-gateway-fips package becomes kong-ai-gateway-fips-140-3 and the <version>-<distro>-fips image becomes <version>-<distro>-fips-140-3. The new package replaces the old one, so an installed FIPS deployment upgrades in place, but pinned package names and image tags must be updated.

Plugin

  • openid-connect:

    • KONG_FIPS=on no longer generates default EdDSA JWKS keys. Use an RSA-PSS or NIST-curve ECDSA key instead.

Feature

Core

  • Added a Distroless Docker image variant for Kong AI Gateway. The Distroless image ships only the gateway binary and its runtime dependencies. It has no shell, no package manager and no general-purpose userland.

  • Added a FIPS 140-3 variant of the Kong AI Gateway Distroless image, published as <version>-distroless-fips-140-3 for amd64.

  • Added a FIPS-compliant variant of the distroless image.

  • ai-mcp-proxy: Added PDK support for effective requests, original request metadata in log, and pre-finalization response status and headers in body_filter.

  • ai-proxy: Added support for TypeSafe.ai Jev decision models, either hosted on TypeSafe or a third-party (via custom upstream URL).

  • Analytics: Added the W3C trace context headers (traceparent, tracestate) and, for Bedrock, SageMaker, and Vertex AI (Gemini), the inference provider’s region (ai[].meta.provider_region) to the request payload.

  • Enabled TLS 1.3 for the FIPS cipher suite and configured FIPS builds to use AES-based TLS 1.3 ciphersuites.

PDK

  • Added a new Metrics PDK (kong.metrics) that enables custom plugins to define and record metrics, with initial support integrated into the opentelemetry plugin.

Plugin

  • AI Proxy:

    • Added the ability to choose whether to output OpenTelemetry span attributes and statistics in OpenInference format, or the previous OpenLLMetry format.

  • ai-mcp-proxy:

    • Added Token Vault support to resolve upstream credentials, allowing the exchange and forwarding of short-lived credentials instead of static ones.

  • ai-plugins:

    • Introduced a new filter execution ordering system so that the shared filters in AI plugins could be re-ordered. This helps bring more flexibility to the execution.

  • ai-proxy-advanced:

    • Added a passthrough route type that forwards the client’s request path and body and the upstream response bytes unchanged, while still applying provider authentication.

    • Added support for the OpenAI and Anthropic Skills API, with the client and the upstream provider using the same format (OpenAI-to-OpenAI, Anthropic-to-Anthropic).

  • ai-prompt-compressor:

    • Added a headroom compression provider, selected with the new config.provider field. This is the stateful Headroom mode: the sidecar keeps the compressed conversation on its own side and indexes that state by session_id, which the plugin derives per conversation.

  • ai-rate-limiting-advanced:

    • Added credential and service match types to policies, so a policy can filter on or partition by the authenticated credential ID, and on the service ID or (with key: name) the service name.

  • opentelemetry:

    • Added the client_certificate config option, so users can present a certificate entity as the mTLS client certificate when an OTLP endpoint uses HTTPS.

Clustering

  • Required identical AI Gateway versions on control planes and data planes before allowing full or incremental configuration synchronization. Data planes with missing versions or different patch or prerelease versions were rejected.

Bugfix

Plugin

  • AI Analytics:

    • Fixed an issue where ai.model_id were absent in Analytics

  • ai-a2a-proxy:

    • Fixed an issue where proxy_config was not applied to the final A2A upstream request.

    • Rejected auth.provider: token_vault at config validation time, since this plugin only implements aws_iam auth.

  • ai-mcp-proxy:

    • Fixed an issue where setting server.tag in conversion-listener mode returned an empty tool list.

    • Fixed an issue where modern MCP mirroring headers were forwarded to legacy upstream servers.

    • Fixed an issue where plugins before ai-mcp-proxy were run twice in conversion-listener mode.

  • ai-proxy-advanced:

    • Added prompt cache control mappings across OpenAI, Anthropic, and Amazon Bedrock chat formats.

    • Fixed a path-duplication bug on Anthropic restful routes (llm/v1/batches, llm/v1/skills) when the model used a custom upstream_url.

    • Fixed an issue where Gemini image generation misattributed modality tokens and omitted cache usage in responses and AI logs.

    • Fixed an issue where gen_ai.request.max_tokens attribute was missing from OpenTelemetry and debug-session spans.

    • Fixed an issue where Anthropic-compatible streaming providers like Baseten reporting prompt usage only in message_delta lost prompt and cached tokens.

    • Fixed an issue where Anthropic thinking.type = "disabled" was overridden by output_config.effort during cross-format conversion, keeping reasoning enabled.

    • Fixed an issue where an Anthropic thinking block (adaptive, enabled, or disabled) was always converted into a reasoning_effort request parameter and forwarded to the target model, even when that model did not support reasoning. OpenAI, Mistral, and Cohere targets without reasoning support now have the gateway-derived value omitted, while an explicit reasoning_effort sent by the client is still forwarded as before.

    • Fixed an issue where Bedrock and SageMaker assume-role credentials were not prewarmed and could trigger STS calls during balancer-phase retries.

    • Fixed an issue where the consistent-hashing balancer routed requests with the same hash_on_header value to different targets when multiple workers were running.

    • Fixed an issue where repeated cumulative usage in streaming responses inflated token counts.

    • Fixed an issue where xAI reasoning tokens were not folded when visible completion tokens exceeded reasoning tokens.

    • Fixed an issue where Anthropic streaming responses and Bedrock native-format responses did not report web search and web fetch tool usage counts in observability metrics.

    • Fixed an issue where Anthropic output_config.effort was ignored during cross-format conversion (including Bedrock additionalModelRequestFields), and added support for preserving max effort.

  • acme:

    • Fixed an issue where an expired fail_backoff key would keep blocking certificate issuance. Retries are now rejected only while the backoff timestamp is still in the future.

  • ai-aws-guardrails:

    • Fixed handling of OpenAI-compatible multimodal requests with image content.

  • ai-prompt-compressor:

    • Fixed an issue where, on routes where the client and the upstream use the same format, the compressor sent the original request instead of the one modified by ai-sanitizer, masking guardrails, ai-rag-injector or ai-request-transformer.

  • ai-response-transformer:

    • Fixed an issue where a chained upstream’s raw, provider-native response was sent to introspection and returned to clients unnormalized.

  • ai-sanitizer:

    • Fixed an issue where ai-sanitizer only sanitized the first Gemini response candidate, leaving additional candidates unsanitized when candidateCount was greater than one.

    • Fixed an issue where ai-sanitizer failed to sanitize Cohere requests that contained multimodal message content, such as image blocks alongside text.

  • datakit:

    • Fixed an issue where the exit node sent a string body without a correct Content-Type. It now uses the content type that the source node declares, for example application/xml for json_to_xml output, and defaults to application/octet-stream in the request phase.

  • graphql-rate-limiting-advanced:

    • Fixed an issue where a longer fixed window was incorrectly marked as exhausted if a shorter window denied a request.

  • jwt:

    • Fixed an issue where the plugin failed to decode standard Base64 secrets that contained + or / when config.secret_is_base64 was enabled.

  • opa:

    • Fixed an issue where Kong Gateway did not send an SNI to the OPA server. It now sends opa_host as the SNI when that is a hostname; with ssl_verify = true, the certificate must include opa_host.

  • openid-connect:

    • Fixed an issue where disabling verify_signature gave no warning, even though it also skips signature verification for bearer tokens from clients. The plugin now logs a warning when verify_signature is false and auth_methods includes bearer. It logs a notice instead when bearer is not enabled.

    • Fixed an issue where a callback with an invalid JARM response returned a 500 error; the plugin now rejects such callbacks with a 401 authentication failure.

  • opentelemetry:

    • Fixed an issue where metrics were rarely (or never) exported on data planes whose configuration is rebuilt more frequently than the metrics push_interval. The periodic metrics push timer is no longer cancelled and recreated on every configuration rebuild, so its push_interval countdown is preserved.

  • rate-limiting-advanced:

    • Fixed an issue where a longer fixed window was incorrectly marked as exhausted if a shorter window denied a request.

  • request-callout:

    • Fixed an issue where a callout that failed and then succeeded on retry still returned the earlier failed response. The plugin now returns the response from the successful retry.

  • service-protection:

    • Fixed an issue where a longer fixed window was incorrectly marked as exhausted if a shorter window denied a request.

Core

  • AI Gateway: Fixed the RPM Obsoletes metadata of the AI Gateway packages. kong-ai-gateway and kong-ai-gateway-fips-140-3 are parallel builds, not successors of each other, so they now obsolete each other only below the version being built. Without the bound an installed kong-ai-gateway had two mutually conflicting successors and dnf upgrade could not resolve an upgrade. The retired kong-ai-gateway-fips keeps a single unversioned successor, kong-ai-gateway-fips-140-3, so an existing FIPS deployment upgrades in place. The deb packages do not change.

  • Analytics: splunk reports now carry ai_a2a_reqs and ai_total_reqs fields, counting A2A requests and all AI Gateway traffic (inference, MCP, and A2A) respectively.

  • Balancer: Fixed an issue where a plugin could not override the upstream target when the Gateway Service host was a literal IP address. The balancer kept the IP that it derived from the Service, so the request went to the Service host and the override was silently lost. This affected kong.service.set_target and every plugin that replaces ngx.ctx.balancer_data.host, for example route-transformer-advanced and canary.

  • Fixed an issue where a malformed JSON document could corrupt the shared JSON decoder of a worker. Later requests that decoded valid JSON on the same worker then failed until the node restarted. The JWT plugin returned Bad token; invalid JWT payload for known-good tokens.

  • Fixed an issue where an in-flight request that completed after its upstream target became unavailable could restore the target to the latency balancer and route subsequent requests to it.

  • Fixed an issue where custom_validator in schema definitions forwarded the enclosing schema as a second argument. This changed the call signature and could affect custom validators that declare more than one parameter. The single-argument call signature is restored.

  • Fixed an issue where the SameSite attribute of the Set-Cookie header was incorrectly set to Same-Site by session-affinity (sticky-session) balancing.

  • On shell-free (distroless) images, the pluginserver query and start commands now expand $KONG_PREFIX / ${KONG_PREFIX} from the environment and fail fast with a clear error on an unterminated quote or ${, instead of passing the literal $KONG_PREFIX to the pluginserver or building a best-effort argv.

Clustering

  • Fixed an issue where older data planes stopped receiving configuration updates when sync v2 was unavailable by restoring automatic sync v1 fallback.

CLI Command

  • Hardened stale nginx PID detection to survive PID reuse. On Linux, kong start now confirms a live PID actually belongs to nginx by reading /proc/<pid>/cmdline, so a reused PID from an ungraceful exit no longer blocks a restart.

Dependency

Core

  • Bumped kong-pgmoon from 2.3.4.0 to 2.3.5.0, which fixes SCRAM channel binding failing to connect to a PostgreSQL server whose certificate is signed with SHA-512.

  • Bumped lua-resty-jq from 0.2.0 to 0.2.2.

  • Bumped lua-resty-simdjson from 1.2.0 to 1.3.1 (updates simdjson to 4.6.11).

  • Bumped OpenSSL FIPS provider from 3.0.9 to 3.1.2 for FIPS 140-3 validation

Performance

Plugin

  • ai-mcp-proxy:

    • Cached Token Vault credential exchanges using Redis to reduce repeated upstream credential exchanges per request.

  • oas-validation:

    • Improved plugin performance by removing unnecessary hash computation. Enabling the plugin now adds much less latency and CPU overhead to requests.

2.1.0

Release date 2026/09/24

Breaking Change

Core

  • Conditional Plugin Execution: The plugin condition field now uses a new standardized expression language with broader operator and type support, aligned across UI, Control Plane, and Data Plane. Refer to the official documentation for the full syntax, supported variables, operators, and types.

  • Workspace names that shadow existing top-level admin API routes are no longer allowed. If you have an existing workspace name that conflicts with a admin route, it may be impacted. Check GET /endpoints to review the reserved first-level segments.

Feature

Plugin

  • ai-mcp-proxy:

    • Added config.server.cache for the ttlMs and cacheScope hints Kong emits on tools/list and server/discover, and config.server.allowed_versions to narrow the MCP protocol revisions this server accepts. allowed_versions is unset by default, meaning every revision Kong implements; when set, server/discover advertises exactly that list and a request declaring anything else is rejected with -32022. A public cache scope is rejected on a configuration whose tool list is ACL-filtered per subject.

    • Added propagation of the inbound W3C trace context to MCP tool call requests.

    • Added support for speaking MCP protocol revision 2026-07-28 to upstream MCP servers. Added config.server.upstream_protocol_version to select the revision Kong speaks upstream.

    • Added support for MCP protocol revision 2026-07-28. Kong answers a client at the revision it declares, resolved per request from the MCP-Protocol-Version header and the _meta protocol version. The revision brings server/discover, subscriptions/listen, resultType, ttlMs/cacheScope cache hints, HTTP status codes on protocol errors, header mirroring validation against the request body, and no sessions or standalone GET stream. 2025-03-26, 2025-06-18 and 2025-11-25 clients are unaffected.

  • ai-proxy-advanced:

    • Added handling for JSON null metadata fields in native-format analytics.

    • Added per-modality usage cost for text, audio, image and video tokens, configured through the pricing catalog or the new model.options *_cost_list fields.

  • acl:

    • Added allow_when/deny_when fields that evaluate CEL boolean expressions against the request context to provide more expressive access‑control policies.

  • openid-connect:

    • Added support for reading bearer tokens from a custom header.

CLI Command

  • Added a kong workspace CLI command. kong workspace rename OLD_NAME NEW_NAME renames a workspace (cascading the new name to RBAC role-endpoint permissions), and kong workspace detect-admin-endpoint-conflicts lists workspaces whose names collide with admin API URL prefixes.

Core

  • Analytics: MCP analytics now carry protocol_version, the revision the client declared, and tool_list_cache_hit, whether a tools/list was answered from the gateway tool cache, alongside the JSON-RPC error code per RPC. mcp_session_id is empty from revision 2026-07-28 on, which mints no session.

  • Conditional Plugin Execution: Added support for conditional plugin execution through a new condition field in the plugin configuration. Plugins can now be conditionally skipped at runtime based on request attributes such as HTTP method, path, headers, query parameters, consumer, route, service, and network information. The condition expression uses the ATC expression language and is evaluated once during the collecting phase; skipped plugins are excluded from all downstream phases automatically.

  • Konnect analytics now propagates a custom string key/value map on the request payload (RequestMetadata.custom). Plugins can populate it with kong.log.set_serialize_value("custom.<key>", "<value>").

  • Plugin conditions: Plugin condition expressions are now type-checked at configuration time. Expressions whose result is provably non-boolean (for example 1 + 1, "foo", or size(x)) are rejected when the plugin is configured, instead of failing at request time. Boolean and statically-unknown expressions are accepted as before.

  • Supported plugin cloning. You can now clone existing bundled plugins, give them a new name and optionally specify a priority. This is useful when you want to have multiple instances of a plugin with different configurations running in the same request. Currently, the cloneable plugins are: openid-connect, acl, pre-function, post-function, request-transformer-advanced, request-transformer, response-transformer-advanced, response-transformer, key-auth, file-log, http-log, tcp-log, ip-restriction, route-by-header, opa, datakit.

  • Supported plugin streaming. You can now stream custom plugins dynamically. Refer to the documentation for more details on how to use this feature.

Clustering

  • Required identical AI Gateway versions on control planes and data planes before allowing full or incremental configuration synchronization. Data planes with missing versions or different patch or prerelease versions were rejected.

Bugfix

Plugin

  • AI Analytics:

    • Fixed an issue where ai.model_id were absent in Analytics.

  • ai-a2a-proxy:

    • Fixed an issue where non-A2A and oversized requests could reach upstream services without configured AWS authentication.

  • ai-mcp-proxy:

    • A request declaring a protocol revision Kong does not serve is now answered with HTTP 400 and JSON-RPC error -32022, listing the revisions Kong serves, instead of being served silently at a different revision.

    • Fixed an issue where AWS request signing did not read file-backed request bodies within the configured size limit.

    • W3C trace context sent in params._meta (traceparent, tracestate, baggage) now reaches the upstream as headers on the request Kong actually sends, in every mode. A value containing a control character is dropped instead of being written to a header.

  • ai-mcp-proxy/ai-a2a-proxy:

    • Fixed an issue where the mcp_server_id and agent_id fields were missing from the AI analytics telemetry payload unless the plugin was scoped to a gateway service.

  • ai-proxy:

    • Fixed an issue where parallel tool call results were sent to Bedrock in the wrong order, causing a 400 “Expected toolResult blocks” error.

  • ai-proxy-advanced:

    • Added video token metrics to the usage token details.

    • Fixed an issue where a Gemini response with JSON null token counts in usageMetadata caused an HTTP 500 on non-streaming chat requests and an empty HTTP 200 response on streaming chat requests.

    • Fixed an issue where adapted Anthropic responses omitted cache and reasoning usage details.

    • Fixed an issue where Claude Sonnet 5 and Opus 5 rejected OpenAI-format reasoning requests on Bedrock due to an unsupported thinking.type value.

    • Fixed an issue where provider-reported reasoning-token details were not correctly reported.

    • Fixed an issue where some providers would not properly respect the stream_options: include_usage flag, and return out-of-order SSE stream end frames.

    • Fixed an issue where xAI reasoning tokens were not folded into completion_tokens, breaking the OpenAI usage identity.

    • Fixed an issue where the consistent-hashing balancer routed requests with the same hash_on_header value to different targets when multiple workers were running.

    • Fixed empty and statistics-only SSE streams. Empty Anthropic responses no longer emit an unmatched content_block_stop, and log_statistics-only streams now track streamed response length for completion-token estimation.

    • Fixed multipart video generation requests to OpenAI-compatible providers leaking internal multipart metadata into JSON upstream bodies, which caused the requests to fail.

  • ai-sanitizer:

    • Fixed an issue where ai-sanitizer returned 500 status code when the provider was using another model, such as a claude-* model served by the gemini provider on Vertex AI.

  • datakit:

    • Fixed an issue where the exit node sent a string body without a correct Content-Type. It now uses the content type that the source node declares, for example application/xml for json_to_xml output, and defaults to application/octet-stream in the request phase.

    • Fixed the body handling of the exit and service_request nodes when the body input is a JSON null. The nodes sent the internal Lua value as the string “userdata: NULL”. The exit node now sends no body. The service_request node now keeps the upstream request body.

  • jwt:

    • Fixed an issue where the plugin failed to decode standard Base64 secrets that contained + or / when config.secret_is_base64 was enabled.

  • oas-validation:

    • Fixed an issue where the OAS specification cache failed to distinguish between different values of api_spec_encoded for the same specification content.

  • ai-semantic-prompt-guard, ai-semantic-response-guard, ai-proxy-advanced, ai-semantic-cache: Fixed vector database indexes and cached embedding records being left behind when a plugin instance was deleted.

  • kafka-consume, confluent-consume: Fixed an issue where the consumer group subscription declared protocol version 1 but carried the version 0 layout, so a topic-rewriting proxy rejected it.

  • prometheus, opentelemetry: Fixed an issue where prompt-cache token usage was missing from observability: added per-TTL cache-write token types (e.g. cache_write_1h_tokens, cache_write_5m_tokens) to the Prometheus ai_llm_tokens_total and OpenTelemetry gen_ai.client.token.usage metrics, and gen_ai.usage.cache_creation.input_tokens / gen_ai.usage.cache_read.input_tokens to OpenTelemetry traces and access logs.

Core

  • Fixed an issue where an in-flight request that completed after its upstream target became unavailable could restore the target to the latency balancer and route subsequent requests to it.

  • Fixed an issue where custom_validator in schema definitions forwarded the enclosing schema as a second argument. This changed the call signature and could affect custom validators that declare more than one parameter. The single-argument call signature is restored.

  • Fixed an issue where streamed custom plugins were not recognized when configured in the ordering field of plugins on Data Planes.

  • Fixed an issue where the priority field of a cloned plugin could not be updated while it was still referenced by plugin instances.

Clustering

  • Fixed an issue where older data planes stopped receiving configuration updates when sync v2 was unavailable by restoring automatic sync v1 fallback.

  • Fixed an issue where paginated full sync could keep stale pending foreign-key references across a failed resume, causing spurious validation failures.

  • Fixed an issue where self-managed control planes rejected updates to a streamed plugin’s handler while the plugin was referenced by a plugin instance.

  • Fixed an issue where updates to a streamed plugin’s definition (e.g. handler), or to a cloned plugin’s definition (e.g. priority), were not applied on data planes when no referencing plugin instances changed. The data plane now rebuilds its plugins iterator so the updated definition takes effect, instead of continuing to execute the previous version.

Dependency

Core

  • Uses now cel-cpp 0.15.0 instead of cel-rust 0.11.6 for Common Expression Language backend.

Performance

Clustering

  • Improved the configuration synchronization latency when incremental sync enabled under some circumstances.

2.0.3

Release date 2026/08/31

Feature

Plugin

  • AI Proxy:

    • Added support for GCP Workload Identity Federation using AWS IAM to authenticate with GCP Vertex models, enabling Kong to access Google Cloud Vertex models using AWS credentials, and without managing long-lived credentials.

  • ai-mcp-oauth2:

    • Add 401 JSON body like Kong key-auth plugin.

Core

  • Konnect Only. Analytics: Added the OpenID Connect plugin’s credential_claim identity to Analytics events, so requests are attributed under the auth dimension in Konnect without a Kong Consumer.

Bugfix

Plugin

  • ai guardrail plugins:

    • Fixed an issue where a JSON null field in a response from the third-party service made the plugin return 500.

  • ai-mcp-proxy:

    • Fixed an issue where gzip-compressed and malformed MCP request bodies were incorrectly rejected with HTTP 413.

  • ai-model-selector:

    • Fixed an issue where a model extracted from a body_path other than model left an unrecognized field in the request body and no standard model field, causing the upstream LLM provider to reject the request.

  • ai-proxy and ai-proxy-advanced:

    • Fixed a 500 crash on Bedrock Converser requests when a cachePoint content block immediately followed a toolResult block in the same message.

  • ai-proxy-advanced:

    • Fixed an issue where Gemini non-streaming responses that reached max_tokens reported a stop finish reason instead of length.

    • Fixed an issue where modern OpenAI output token limits were discarded or overridden during request normalization.

    • Fixed an issue where Gemini and Anthropic token usage metadata was lost in streaming responses.

  • ai-aws-guardrails:

    • Fixed a issue where the request to aws guardrail service could return 500 error duo to a failure in request signature.

  • ai-rag-injector:

    • Fixed an issue where a chunk lookup against an Amazon MemoryDB vector store failed with unexpected search response: no value found in result set. The MemoryDB driver now reads the FT.SEARCH fields by name instead of by position, treats the vector distance as optional, and requests as many rows as fetch_chunks_count asks for.

  • jwt:

    • Fixed an issue where the plugin failed to decode standard Base64 secrets that contained + or / when config.secret_is_base64 was enabled.

  • openid-connect:

    • Fixed an issue where a request that authenticated but failed an authorization check (such as scopes_required) was not attributed to its consumer, consumer groups, or Kong Identity principal in logs and analytics.

  • opentelemetry:

    • Reduced per-request overhead of resource_attributes by skipping template rendering entirely for static (non-templated) fields.

  • kafka-upstream, kafka-log, confluent, kafka-consume, confluent-consume: Fixed a permanent failure that started when a broker answered an ApiVersions request with an error code, an empty API key set, or a malformed frame. The client cached that result and every later request failed with ProduceRequest API version not negotiated; broker may be unreachable until the node restarted, even after the broker recovered. The client now refuses an unusable ApiVersions response, keeps the previous negotiation, and retries the negotiation for a topic whose metadata is already cached.

Core

  • Fixed an issue where an in-flight request that completed after its upstream target became unavailable could restore the target to the latency balancer and route subsequent requests to it.

  • Fixed an issue where immediate cosocket SSL handshake errors could be reported as subsequent socket write failures.

  • Konnect Only. Debugger: Fixed an issue where deep copy of the worker logs stack overflowed by replacing deep_copy with cycle_aware_deep_copy.

Clustering

  • Fixed an issue where hybrid data planes retried unrecoverable CP connection failures and ignored Retry-After delays on 429/503 responses, preventing avoidable reconnect storms.

Performance

Clustering

  • Clustering: Improved data plane stability when applying a large configuration sync from the control plane, reducing the chance of a worker restart while the update is being processed.

  • Optimized performance of clustering payload handling for version compatibility. This reduces CPU usage and transient memory pressure.

Admin API

  • Improved the performance of listing workspaces with ?counter=true. Entity counts for the whole page are now retrieved with one grouped query per entity type instead of a separate count query per entity type per workspace, while keeping the counts exact.

Plugin

  • opentelemetry:

    • Reduced memory usage of the access log queue by deferring OTLP payload conversion to export time.

2.0.2

Release date 2026/08/18

Feature

Plugin

  • ai-model-selector:

    • Replaced the single source config field with an ordered sources array so the plugin can try multiple sources and use the model from the first one that yields a value.

  • ai-proxy-advanced:

    • Added filtering for empty Anthropic thinking blocks in request history.

    • usage cost now supports per-cache-TTL cache-write pricing, context-window pricing factors, and service-tier pricing factors via model.options (cache_write_cost_list, context_window_factor, service_tier_factor). Anthropic (cache TTL + service tier), OpenAI, and Gemini service tiers are collected from upstream responses.

  • basic-auth:

    • Added support for authentication via the Kong Identity service with caching, consumer compatibility and basic analytics.

  • key-auth:

    • Added support for authentication via the Kong Identity service with caching, consumer compatibility and basic analytics.

  • openid-connect:

    • Added support for Kong Identity principal lookup after token verification, enabling consumer and consumer group mapping from Kong Identity with caching and miss-behavior control.

  • request-transformer-advanced:

    • Added a max_request_body_size configuration option that lets the plugin apply body transformations to request bodies that Nginx buffers to a temporary file (bodies larger than client_body_buffer_size). Previously such bodies were never read, so the configured body transformations were silently skipped and the original body was proxied upstream. The option defaults to -1, which preserves that behavior; set it to a positive byte limit to opt in (or 0 for unlimited, still bounded by Nginx’s client_max_body_size).

  • ace Added Kong Identity principal fetch after authentication, enabling consumer and consumer group mapping from the Konnect identity service.

  • ai-aws-guardrails, ai-azure-content-safety: Added last_message as new optional value for config.text_source.

Bugfix

Plugin

  • ai guardrail plugins:

    • Fixed issues where ai-sanitizer plugin and ai-aws-guardrail plugin can not work properly with claude code.

  • ai-proxy, ai-proxy-advanced:

    • Fixed an issue where Gemini streaming analytics could over-report token usage. Gemini repeats its usage metadata on every stream chunk as a running request-level total, but the counts were summed across chunks instead of replaced, multiplying the reported usage by the number of chunks. Reasoning tokens were affected the most, because Gemini reports them in full on the first chunk and does not change them afterwards. A chunk that carried no usage metadata also reset the recorded output token count to zero.

  • ai-proxy-advanced:

    • Fixed a bug where a response from Anthropic could fail to decode when the client’s request advertised Brotli support, since Anthropic’s edge may then return a Brotli-encoded body that Kong’s response decompression – which only understands gzip – can’t inflate. Kong now normalizes the outgoing Accept-Encoding request header before proxying to any LLM provider: gzip, identity when the client accepts gzip, and identity otherwise, so a client that cannot inflate gzip is never handed a compressed body on the paths that pass the upstream response through untouched.

    • Fixed an issue where Gemini’s tool-use token usage (billed when server-side tools such as grounding, code execution, or URL context are used) was not counted in the reported token usage or cost.

    • Fixed an issue where signed Anthropic thinking blocks with omitted text were removed from request history.

    • Fixed an issue where token usage for Anthropic requests using the compaction beta was undercounted. Kong now sums usage.iterations when present instead of relying solely on the top-level input_tokens / output_tokens, which only reflect non-compaction iterations.

    • Fixed an issue where Anthropic-format requests could report zero usage when it targets Bedrock’s own model via conversion.

    • Fixed custom Azure upstream URLs to preserve query parameters and apply the configured api-version when needed.

    • Fixed Gemini image generation token accounting to include reasoning tokens (thoughtsTokenCount) in output_tokens and report them as output_tokens_details.reasoning_tokens. Previously the reasoning tokens were dropped, which under-reported output tokens, analytics, and calculated cost for thinking-enabled image models.

    • Fixed Gemini realtime (Live API) token accounting to include reasoning tokens (thoughtsTokenCount) in the reported output tokens, and to report them as output_tokens_reasoning_tokens_count. Previously the reasoning tokens were dropped, which under-reported the output and total token counts of thinking-enabled realtime sessions.

    • Fixed token usage statistics being silently dropped for audio/v1/audio/transcriptions requests. Token-billed OpenAI/Azure transcription models (e.g. gpt-4o-transcribe) now report input_tokens/output_tokens/total_tokens and the text_tokens/audio_tokens breakdown in the analytics log when logging.log_statistics is enabled.

  • ai-aws-guardrails:

    • Fixed an issue where the plugin ignored the configured text_source option when allow_masking was enabled, masking the entire raw request body instead.

  • ai-prompt-guard:

    • Fixed an issue where the plugin failed to inspect requests sent in a provider’s native format instead of the OpenAI-compatible format.

  • ai-sanitizer:

    • Fixed an issue in OpenAI chat response extraction where tool-call-only choices (without string message.content) could be treated as extractable text.

  • jwt-signer:

    • Fixed an issue where the plugin failed to set the signed access token header on the upstream request, when access_token_upstream_header (or original_access_token_upstream_header) had the same name as channel_token_request_header.

  • opentelemetry:

    • Fixed gen_ai.client.token.usage never reporting token counts for non-text-generation AI categories (image, video, realtime, and audio transcription generation), because it only read the legacy prompt_tokens/prompt_tokens_details analytics log fields instead of falling back to the input_tokens/input_tokens_details fields those categories are logged under.

  • prometheus:

    • Fixed kong_ai_llm_tokens_total never reporting token counts for non-text-generation AI categories (image, video, realtime, and audio transcription generation), because it only read the legacy prompt_tokens/prompt_tokens_details analytics log fields instead of falling back to the input_tokens/input_tokens_details fields those categories are logged under.

  • Fixed an issue where AI Gateway guardrail plugins’ request extractors did not correctly honor the configured text_source option.

  • Fixed an issue where AI Gateway guardrail plugins failed to extract prompt text from OpenAI-native requests using the input_text content type.

Performance

Core

  • Core: Reduced config update latency on hybrid data planes by notifying the stream subsystem before applying the update in the HTTP subsystem, so both rebuild in parallel.

2.0.1

Release date 2026/07/29

Feature

Plugin

  • ai-guardrails:

    • Added a rejection_mode configuration field (none, stealth, verbose) to control how blocked requests and responses are communicated to the client.

    • Added a shared continue_on_detection field to all guardrails to allow for non-blocking guardrail triggers.

  • ai-mcp-proxy:

    • Added JSON Schema draft 2020-12 support for input schema validation.

  • ai-model-selector:

    • Added support for extracting model aliases from request paths.

  • ai-proxy:

    • Added support for AWS SageMaker as a new AI provider.

  • ai-proxy-advanced:

    • Added a configurable pricing catalog that computes LLM usage.cost from per-(provider, model) token rates, including cache read/write rates.

    • Added OpenAI-format chat reasoning request and response mappings across supported AI providers.

    • Added support for reporting prompt cache token metrics in OpenAI-compatible format.

  • ai-proxy/ai-proxy-advanced:

    • allowed ai-proxy/ai-proxy-advanced plugins to be scoped with the ai-model entity.

  • ai-custom-guardrail:

    • Added the log_blocked_content field to allow logging prompts and responses blocked by the guardrail, and updated the continue_on_detection description to mention that block data is visible via Kong logs.

  • ai-mcp-oauth2:

    • Added expected_audience to configure the expected token audience independently from resource.

  • ai-nvidia-nemo-guardrail:

    • Added new plugin to check LLM request/response content against NVIDIA NeMo Guardrails.

  • ai-semantic-response-guard:

    • added the response_buffer_size option to support guarding streaming responses.

Bugfix

Plugin

  • ai:

    • Fixed the Redis vector database index being dropped and recreated on every request when metadata fields were enabled.

  • ai-mcp-proxy:

    • Fixed an issue where the ACL decision cache has collision.

    • Fixed an issue where the MCP initialize response reported the wrong version.

    • Fixed an issue where non-managed MCP listeners advertised unsupported SSE streams.

  • ai-model-selector:

    • Fixed model alias extraction from the request body to support multipart/form-data requests in addition to JSON.

  • ai-proxy:

    • Fixed a confusing error message when a Bedrock native-format request uses an invalid ARN model path.

    • Fixed an issue where Bedrock driver could not handle OpenAI content_parts format in system messages, user/assistant messages, and array-typed tool results.

    • Fixed OpenAI-compatible completions requests with string array prompts.

  • ai-proxy-advanced:

    • Fixed a Lua crash in declarative schema validation when a Bedrock target had no auth block (e.g. IRSA / role-only configs), which made the entire declarative config unloadable.

    • Fixed an issue where the configured auth header (e.g. api-key, x-api-key) leaked the provider credential into access logs instead of being redacted.

    • Fixed an issue where the OpenAI developer role was not treated as a system prompt when transforming requests to Bedrock and Gemini providers.

    • Fixed an issue where we didn’t handle SSE event headers correctly.

    • Fixed an issue where balancer retries could continue reusing the last selected target after all priority peers were exhausted.

    • Fixed an issue where Gemini chat requests rejected OpenAI input_audio and video_url content parts.

    • Fixed an issue where native Bedrock request parsing dropped string message content in kong internal format, causing invalid data used by other plugins.

    • Fixed an issue where OpenAI content array was sent as bedrock/gemini system message which only support string.

    • Fixed an issue where OpenAI, Azure, and Gemini realtime upstream URLs using wss were downgraded to ws.

    • Fixed an issue where requests to a Gemini or Anthropic provider could fail with 401 when using the OpenAI format with a custom auth header (e.g. x-goog-api-key, x-api-key) or query-parameter auth. The original Authorization header sent by the OpenAI SDK is now cleared so it does not conflict with the configured credential.

    • Fixed an issue where semantic routing failed with “found target with ID but it's not mapped to a target" when multiple workers or nodes raced to create the vectorDB index: "Index already exists" errors from concurrent index creation are now treated as success in the Redis, Valkey, and MemoryDB strategies, so every worker completes initialization and builds its local target mapping.

    • Fixed an issue where the context-management-2025-06-27 Anthropic beta flag was dropped when routing to a Bedrock target, causing Bedrock to reject requests carrying context_management with a 400 error on failover from an Anthropic target.

    • Fixed an issue where a malformed RSA private key in a Gemini provider’s service_account_json caused crash with SIGSEGV

    • Fixed Gemini service-account auth probing metadata.google.internal when a service account JSON was configured.

    • Fixed reuse of an expired Azure managed-identity token causing intermittent upstream 401s.

  • acl:

    • Fixed an issue where the ACL plugin returned 403 for authenticated credentials with no consumer and no groups (e.g. ACE / Dev Portal application credentials) when configured with a deny list. A group-less identity cannot match a deny list and should be allowed through.

    • Fixed an issue where the ACL plugin ignored a consumer’s ACL groups when include_consumer_groups was enabled and the consumer belonged to an enterprise consumer group, causing incorrect 403 denials (with allow) or incorrect access (with deny) for authenticated and anonymous consumers. The plugin now falls back to the consumer’s ACL groups when no third-party authenticated groups are present.

  • ai-mcp-oauth2:

    • Fixed a JWT security vulnerability.

  • ai-prompt-compressor:

    • Fixed an issue where native Anthropic requests were forwarded in OpenAI format after prompt compression.

  • ai-rag-injector:

    • Fixed percent sequences in injected prompt text.

  • ai-rate-limiting-advanced:

    • Fixed token counters being decremented on ai-semantic-cache hits.

  • ai-request-transformer:

    • Fixed an issue where requests transformed before ai-proxy-advanced skipped AI response normalization.

  • ai-semantic-cache:

    • Fixed cache hits ignoring llm_format, which returned OpenAI-shaped responses to native-format (e.g. Anthropic) clients. Cached responses are now converted to the client’s llm_format for both streaming and non-streaming requests.

  • datakit:

    • Fixed CVE-2026-14916 by binding the JWT algorithm to the verification key’s type in the jwt_verify node, so a symmetric algorithm can no longer be verified against an asymmetric key, preventing an authentication bypass where an attacker forges tokens by using the public key as an HMAC shared secret, under configurations where the node uses a PEM key string and allowed_algorithms is unset (any algorithm accepted).

  • ip-restriction:

    • Fixed an issue where warn log messages did not include the plugin name, making it difficult to identify the source of blocked requests.

  • jwt-signer:

    • Fixed an issue where the plugin failed to load configuration and errored when null or nil values were passed to access_token_keyset or channel_token_keyset.

  • openid-connect:

    • Fixed an issue where switching clients within a single request (via client_arg) failed to use the matching client’s JWK.

    • Fixed an issue where consumer groups could not be resolved when the groups claim was only available from the user info endpoint, which resulted in a 500 error.

  • opentelemetry:

    • Fixed an issue where OpenTelemetry metrics only supported integer values by adding support for floating-point values in the OTLP exporter.

  • prometheus:

    • A histogram’s series (its buckets, _count and _sum) is now read as one atomic group during a scrape. Previously the exporter yielded and flushed output between individual samples, so an observation landing mid-scrape could bump _count/_sum after the buckets were already read, producing an internally inconsistent histogram in a single response.

    • Histogram series are now emitted grouped by label set (all buckets in ascending le order, then _count, then _sum) so the output follows the OpenMetrics interleaving rule. Previously samples were grouped by suffix, which caused strict consumers such as Google Managed Prometheus to drop points.

  • saml:

    • Fixed an unreachable code path that prevented the plugin from falling back to validating the samlp:Response-level signature when the saml:Assertion was not individually signed. With validate_assertion_signature enabled, an unsigned assertion inside a signed Response is now accepted (and a Response with no valid signature on either the assertion or the root is rejected), so IdPs that sign the Response envelope rather than the assertion no longer require signature validation to be turned off. Relates to the Response-level fallback described in CVE-2026-14917.

  • kafka-upstream, kafka-log, confluent, kafka-consume, confluent-consume: Fixed a crash that occurred when a plugin set schema_registry.confluent.authentication.mode to oauth2 but omitted the oauth2_client record. Previously this returned a 500 from the Admin API during validation, and once saved it also crashed at runtime while fetching the schema registry OAuth2 token. The absent oauth2_client is now handled and falls back to the default authentication method, so a configuration without oauth2_client works end-to-end.

Clustering

  • Clustering: The /status/sync endpoint now returns an empty stat object instead of null before the first sync has run, keeping the response shape consistent.

PDK

  • Fixed an issue on HTTP/2 connections where the request method and query were reported incorrectly, or could raise an error, when a request hit the error handler such as a 502 from an upstream connection reset.

Core

  • Fixed an issue where dynamic plugin ordering (ordering) leaked into the downstream phases (header_filter, body_filter, response, log). Dynamic ordering now again applies only to the access phase; downstream phases execute in static plugin-priority order.

  • Konnect Only. Debugger: Fixed an error that could occur under sustained HTTPS traffic.

  • Konnect Only. Debugger: Fixed an issue where the default key name apikey of the key-auth and key-auth-enc plugins was not redacted when captured by the Debugger. This has been resolved by adding apikey to the list of default redacted headers.

  • Konnect Only. Debugger: Fixed an issue where the response body of the external HTTP call httpc:request() were not captured.

  • Konnect Only. Debugger: Fixed an issue where instrumenting third-party HTTP client calls produced two indistinguishable kong.io.http.request spans for httpc:request_uri(). It is now traced under a dedicated kong.io.http.request_uri span, and the connect, DNS, and internal request spans it triggers are nested beneath it. The kong.io.http.request span now also reports the http.url attribute.

  • PostgreSQL OAuth auth: Refined the token cache key to be derived only from the fields that identify an OAuth configuration.

  • router: Fixed an issue where WebSocket routes (ws/wss) returned HTTP 404 when using router_flavor=expressions due to incorrect protocol matching expressions being generated.

Admin API

  • Fixed an issue where resetting the password of an admin or developer failed with “Old password is invalid” when the credential was created before the basic-auth hashing algorithm was upgraded from SHA1 to SHA256 (e.g. a kong_admin user created in an older version and carried over through an upgrade). The old password is now verified with the backward-compatible verification logic that supports both SHA1 and SHA256 hashed passwords.

Dependency

Core

  • Bumped jq from 1.8.1 to 1.8.2.

2.0.0

Release date 2026/07/10

Breaking Change

Plugin

  • ai-plugins:

    • Dropped deprecated schema fields slated for removal in 4.0: http_proxy_host/http_proxy_port/https_proxy_host/https_proxy_port shorthands in ai-llm-as-judge, ai-request-transformer, ai-response-transformer (use proxy_config instead); llm_format, max_request_body_size, and rules.max_request_body_size in ai-semantic-prompt-guard; llm_providers and llm_format in ai-rate-limiting-advanced (policies is now required); model.options.upstream_path and the preserve route_type in the shared LLM schema (use model.options.upstream_url and a concrete route_type instead).

  • Removed the ai-proxy plugin from the bundled plugins list. Use ai-proxy-advanced instead.

Feature

Core

  • Added data plane support for the new AI Gateway control plane. Data planes now use JSON-RPC config sync (sync v2) only and send a validated deployment_type value in the RPC hello payload.

  • AI Gateway: Added a tags field to the ai_models entity, enabling tag-based filtering on the Admin API and the /tags endpoint.

Plugin

  • ai-a2a-proxy:

    • Added A2A v1.0 wire-format support alongside v0.3.

    • now rewrites the URL for all supportedInterfaces in the agent card.

  • ai-guardrail:

    • Add guardrail_triggered field to protobuf analytics payload.

  • ai-mcp-proxy, ai-mcp-oauth2:

    • Introduced forward-proxy support for outbound traffic configuration.

  • ai-mcp-proxy:

    • Added upstream MCP server aggregator support, allowing the plugin to aggregate tools from multiple upstream MCP servers.

    • Updated the default MCP protocol version to 2025-11-25 and added explicit negotiation for older clients.

  • ai-plugins:

    • Added proxy_config support to route plugin HTTP requests through a forward proxy for ai-aws-guardrails, ai-azure-content-safety, ai-custom-guardrail, ai-gcp-model-armor, ai-lakera-guard, ai-sanitizer, ai-prompt-compressor, ai-rag-injector, ai-semantic-cache, ai-semantic-prompt-guard, ai-semantic-response-guard, ai-request-transformer, ai-response-transformer, and ai-llm-as-judge. For ai-request-transformer, ai-response-transformer, and ai-llm-as-judge, the existing top-level http_proxy_host/http_proxy_port/ https_proxy_host/https_proxy_port fields are now deprecated in favor of proxy_config and will be removed in 4.0.

  • ai-proxy:

    • Added OTel span attributes gen_ai.input.messages for audio speech routes and gen_ai.output.messages for audio transcription routes when log_payloads is enabled.

    • Added support for Kimi AI (Moonshot) - a new AI Provider.

    • Added support for Vercel - a new AI Provider.

    • Improved Cohere driver with streaming fixes, tool calling support, and usage tracking.

  • ai-proxy-advanced:

    • Added clarification for handling “Request body too large” errors.

    • Added embeddings support for databricks provider.

    • Added Gemini OpenAI-compatible logprobs mapping on chat completions.

    • Added new ACL matcher type authenticated_groups, which allows propagating claims and roles from other authentication plugins into the AI Proxy Advanced per-model ACLs.

    • Added proxy_config support for forward proxies, including HTTP proxying, Basic authentication, and CONNECT tunneling for HTTPS upstream streaming.

    • Added reasoning fields mapping between Anthropic Messages and the OpenAI internal format.

    • Added support for image analysis request to anthropic/bedrock providers.

    • Added support for anthropic format chat on DeepSeek provider.

  • kimi:

    • Added support for anthropic passthrough on Kimi (Moonshot) provider.

  • llm:

    • Increase max request body size from 1MB to 8MB.

  • vercel:

    • Added support for Anthropic passthrough on Vercel provider.

  • ai-llm-as-judge:

    • Added the opt-in response header to expose the computed judge score to clients. The header name defaults to X-Kong-LLM-Accuracy-Score and is configurable via score_header_name.

  • ai-rate-limiting-advanced:

    • Added calendar window type. Fixed fetch drift in rate limiting windows.

Bugfix

Plugin

  • ai:

    • Fixed an issue where non-200 responses from Bedrock embeddings API were not properly handled, leading to incorrect feedback to the client.

  • ai-a2a-proxy:

    • Reassembled SSE data: events that span multiple body_filter chunks and dropped oversized events safely, preventing sse_last_event type errors that crashed the request on large agent responses.

  • ai-advanced-proxy:

    • Fixed an issue where we never stored response model in streaming path. It always fallback to the request model

  • ai-guardrail:

    • Fixed an issue where latency metrics were missing in the Prometheus output.

  • ai-guardrails:

    • Fixed an issue where multiple guardrail plugins can not function together.

  • ai-mcp-proxy:

    • Fixed an issue where empty JSON arrays in MCP server responses were rewritten to empty objects when the plugin re-encoded the body for ACL list filtering.

    • Fixed an issue where raw non-200 upstream responses triggered misleading MCP body parse warnings.

    • Fixed an issue where WWW-Authenticate: Bearer error=”insufficient_scope” header was not emitted.

    • Fixed an issue where internal unix-socket subrequests lost the original client IP address, causing IP-aware plugins like ip-restriction to reject legitimate requests.

    • Fixed an issue where we didn’t propagate errors on secret_to_jwk calls.

    • Fixed an issue where, since 3.14, call tools converted from API could fail when Kong used a self-signed certificate.

    • Fixed an issue where converted HTTPS MCP tool calls lost the original forwarded port on the internal TLS Unix socket.

    • Fixed an issue where tool id might be repeated.

    • Fixed an issue where we didn’t allow arbitrary tool path in conversion-listener mode.

  • ai-proxy:

    • Fixed an issue where responses with unsupported content types could be returned with an incorrect gzip header.

    • Fixed an issue where log_payloads logged binary data for audio routes, the plugin now logs only the request for speech and only the response for transcription/translation routes.

    • Fixed Azure AI Foundry v1 API support to not require deployment-specific fields when upstream_url is configured.

    • Fixed Gemini driver dropping Gemini-native fields (e.g. cachedContent, safetySettings) and Vertex AI-only fields (e.g. labels) during OpenAI-to-Gemini request transformation.

    • Fixed role matching in message extraction and formatting to be case-insensitive, and added a nil-safety check for messages.

    • Fixed an issue where options.upstream_url was not respected in native mode.

  • ai-proxy-advanced:

    • Fixed an issue where native-format requests using a model alias were incorrectly proxying to upstream model.

    • Fixed an issue where full-sync reconfigure reset balancer state even when plugin configs were unchanged.

    • Fixed an issue where concurrent worker processes could race to initialize the semantic routing vectorDB index, causing “Index already exists” errors on startup or config reload.

    • Fixed an issue where driver transformers returned errors or stream metadata in the wrong return position.

    • Fixed an issue where Vercel Anthropic-format requests did not include the configured model in the upstream body.

    • Fixed an issue where DeepSeek, Vercel, and Kimi Anthropic count-token requests were proxied to chat endpoints.

    • Fixed an issue where OpenAI-compatible Anthropic chat requests only preserved the last system message.

    • Fixed an issue where Anthropic streaming finish reasons are not mapped in OpenAI to Anthropic conversion.

    • Fixed an issue where the Anthropic driver dropped the top_p, top_k, and stop parameters when translating OpenAI-format requests. The driver now also omits temperature, top_p, and top_k for models that reject them (Claude Opus 4.7, Opus 4.8, and Fable), preventing upstream 400 errors.

    • Fixed an issue where output_config was rejected by Vertex AI in Anthropic-native requests.

    • Fixed an issue where cache_creation_input_tokens and cache_read_input_tokens were excluded from prompt token counts, causing undercounting when Anthropic prompt caching was active.

    • Fixed an issue where configured model name may never be used for Azure provider.

    • Fixed an issue where we didn’t support Azure GA version of Realtime API.

    • Fixed an issue where OpenAI-format chat reasoning requests were not mapped consistently across Anthropic, Gemini, Bedrock and more providers.

    • Fixed an issue where system prompt of Claude Code was truncated when doing non-passthrough proxy.

    • Fixed an issue where we didn’t support anthropic format for databricks provider.

    • Fixed an issue where we didn’t use the huggingface’s token usage if available.

    • Fixed an issue where unsafe model names could construct upstream AI provider paths.

    • Fixed an issue where we didn’t collect the model name in OpenAI streaming response from Azure.

    • Fixed an issue where unsafe batch id could cause upstream AI provider path traversal.

    • Fixed an issue where unsafe model name in native format adaptors could cause upstream AI provider path traversal.

    • Fixed an issue where dashscope replied 500 to inference request.

    • Fixed an issue where the xAI driver nested reasoning_effort under reasoning.effort and did not normalize reasoning_content in responses.

    • Fixed an issue where stream_options was forwarded to Databricks upstreams, causing a 400 error since Databricks does not support this field.

    • Fixed an issue where GenAI spans were not finished sometimes.

    • Fixed an issue where observability errors occurred when gpt-image-1.5 returned image token usage details.

    • Fixed an issue where Azure OpenAI authentication failed during load balancer failover due to token refresh attempts in balancer_by_lua. Tokens are now pre-fetched in the access phase.

    • Fixed an issue where cache_control in Bedrock Anthropic format requests using ARN model IDs was lost.

    • Fixed an issue where Bedrock native ConverseStream requests using ARN model IDs could fail while handling streamed responses.

    • Fixed an issue where some Claude Code MCP tools did not correctly handle empty ‘requires’ array.

    • Fixed Azure OpenAI responses support in deployment style.

  • ai-aws-guardrails:

    • Allowed streaming of responses when guarding_mode was set to INPUT with allow_masking enabled.

  • ai-azure-content-safety:

    • Fixed a crash (HTTP 500) when azure_use_managed_identity is enabled.

    • Fixed a bug where the plugin returned HTTP 500 when the LLM response content was empty. Empty content now skips the content-safety check and passes through.

    • Fixed response guard sending the entire JSON response body to Azure Content Safety instead of just the assistant message content, which caused false positives when response metadata fields matched blocklist entries.

  • ai-mcp-oauth2:

    • Fixed an issue where token introspection crashed when client_auth was set to private_key_jwt and client_jwk was provided as a JSON string.

    • Fixed an issue where downstream scope-based ACL evaluation used claims from the inbound token instead of the exchanged token when token exchange was enabled.

    • Fixed an issue where token-related 401 responses omitted the protected resource metadata URL when token was given.

    • Fixed an issue where token exchange actor tokens were not sourced correctly from token_exchange.request, ensuring correct forwarding of actor tokens configured in headers or plugin config.

    • Fixed an issue where token_exchange.cache.enabled = false was ignored and exchanged tokens were still cached because the cache toggle incorrectly read token_exchange.cache.ttl instead of token_exchange.cache.enabled.

  • ai-prompt-decorator:

    • Fixed an issue where the plugin corrupted native-format requests (gemini, anthropic, bedrock, cohere, huggingface) by forwarding an OpenAI-shaped body upstream instead of the original native body shape.

  • ai-rate-limiting-advanced:

    • Emitted the documented X-AI-RateLimit-* headers on 429 responses from policies that partition by provider or model.

    • Fixed an issue where model or provider matchers without partition_by were skipped, causing rate limits to be missed.

    • Fixed policy integration with ai-proxy for single-target and multi-target AI proxy routes.

  • ai-request-transformer:

    • Fixed an issue where runtime template placeholders (e.g. $(uri_captures.*)) in config.llm.model.* fields were not interpolated.

    • Fixed an issue where unsafe model name could cause upstream AI provider path traversal.

  • ai-response-transformer:

    • Fixed an issue where runtime template placeholders (e.g. $(uri_captures.*)) in config.llm.model.* fields were not interpolated.

  • ai-sanitizer:

    • Fixed an issue where the AI sanitizer plugin could not parse the request body correctly.

    • Fixed an issue where the input field for the llm/v1/responses route was not recognized.

  • ai-semantic-prompt-guard:

    • Fixed an issue where max_request_body_size and llm_format were not marked as deprecated.

  • ai-semantic-response-guard:

    • Fixed an issue where the plugin failed to populate Redis with rule documents due to using worker ID gating instead of worker mutex for vectordb operations.

  • openid-connect:

    • Fixed an issue where the redirect_uri parameter was sent to the token endpoint for all grant types. It is now included only for the authorization_code grant, as required by RFC 6749.

  • tcp-log:

    • Fix the race condition where the balancer tries array gets cleared before the asynchronous timer thread can safely encode it.

  • ai-semantic-prompt-guard, ai-semantic-response-guard: Fixed an issue where full-sync reconfigure rebuilt semantic vectorDB indexes even when the plugin configuration was unchanged.

  • Fixed AI Proxy Advanced Gemini failover handling when a target sets upstream_url to ngx.null, avoiding an empty trailing query string in the computed upstream URI.

  • Fixed an issue where when Postgres strategy is used for session storage, concurrent authentication requests could result in race condition and session creation collisions.

Core

  • Applied upstream nginx security patches for CVE-2026-40701, CVE-2026-40460, CVE-2026-42934, CVE-2026-42946, and CVE-2026-42945.

  • Applied upstream nginx security patches for limiting the number of maximum headers (CVE-2026-49975).

  • debugger: Fixed an issue where response body content capture consumed streaming response chunks.

  • Fixed an issue where AI proxy plugins rejected output_cost: 0 in schema validation, preventing users from configuring models that have no output token cost.

  • Fixed an issue where anonymous reports omitted model usage when one provider used multiple models.

  • Fixed an issue where cluster mutex did not timeout immediately with no_wait = true when node level mutex is being held by another worker.

Configuration

  • Increased the default proxy_ssl_verify_depth and lua_ssl_verify_depth from 1 to 5.

Dependency

Core

  • Bumped OpenSSL from 3.5.6 to 3.5.7.

Performance

Plugin

  • ai:

    • Removed unused redis connection when setting up Redis vectordb.

Help us make these docs great!

Kong Developer docs are open source. If you find these useful and want to make them better, contribute today!