litellm/ui
Rohan G 4417bf08ae
feat(guardrails): extend Akto guardrail to responses, MCP tools, attachments and masking (#44343)
* feat(guardrails): extend Akto guardrail to responses, MCP tools, attachments and masking

- post_call now waits for Akto and blocks or masks the reply instead of only logging it
- pre_mcp_call and post_mcp_call check MCP tool arguments and tool results
- attached images, audio and files are sent to Akto's file check
- streamed replies are checked every streaming_sampling_rate chunks
- context_source routes traffic to Akto's endpoint or agentic policies
- tags carry user email, team alias and key alias for attribution

* fix(guardrails): harden Akto MCP detection, keep AGENTIC default, block unmappable masking

- MCP handling trusts the logger's call type, so request body keys can't skip the prompt check
- context_source defaults to AGENTIC, as before
- masking that also hits text we can't write back now blocks
- move tests to tests/unit and rename attachments.py to akto_attachments.py
- regenerate the OpenAPI snapshot and dashboard types

* fix(guardrails): ignore a client-sent response in Akto output checks

- a "response" field sent in the request body is no longer scanned in place of the model's reply
- MCP tool calls in the reply are still checked in that case
- split match or-patterns so CodeQL can follow the bound names
- cover a masked payload that is not JSON and drop unused test imports

* refactor(guardrails): read Akto attachment fields directly instead of pattern captures

* fix(guardrails): check Akto attachments in both messages and input

* fix(guardrails): check every Akto attachment source and keep more prompt text in scope

- check all of a file or image block's sources (file_data, file_url, file_id), since providers pick different ones
- send Anthropic search_result blocks to the file check as text
- keep document title and context, and legacy functions, in the checked request
- take the client IP from the proxy's requester_ip_address before client forwarding headers
- read litellm_params identity only from server-side call details

* fix(guardrails): never drop an Akto attachment the text check removed

- optional metadata (filename, title, format, media type) that isn't a string is ignored instead of failing the block
- an attachment block that still can't be read blocks the request
- search_result text is checked once, as a file, instead of also in the text check

* fix(guardrails): strip only what the Akto file check sends from the text check

- the text check keeps every attachment field except the ones the file check sends
- document title/context and search_result source/title go to the file check as text, since the /v1/messages text check drops them
- accept every image shape LiteLLM forwards (image_url or url, string or object) and check each source
- ignore blocks whose type is not a string instead of failing the file check

* fix(guardrails): keep model-visible text in the Akto text check

- document title/context, text documents and search_result stay in the text check, so no Akto backend skips them
- on /v1/messages the text check reads the messages Anthropic receives, with the guardrail's skip/scan scoping applied
- a client-sent "response" key can only add reply checks, never skip recording or MCP tool-call checks
- a "messages" key on the Responses API can't replace its input in the text check
- the recorded IP comes only from the proxy's requester_ip_address

* fix(guardrails): keep AktoGuardrail positional args backward compatible

* fix(guardrails): keep zero Akto timeouts working and record every stream check

A guardrail_timeout of 0 used to fall back to the default; the new ge=1 made the config invalid, so the proxy dropped the guardrail. Zero settings now fall back to the defaults again.

The end-of-stream check can be skipped when the last sampled check covered the reply, so mid-stream checks now record, like base.

* fix(guardrails): use defaults for non-positive Akto timeouts and sampling rate

A zero or negative guardrail_timeout, file_guardrail_timeout or streaming_sampling_rate used to reach the HTTP call or the stream cadence. They now fall back to the defaults, like unset values.
2026-10-07 15:50:38 -07:00
..
litellm-dashboard feat(guardrails): extend Akto guardrail to responses, MCP tools, attachments and masking (#44343) 2026-10-07 15:50:38 -07:00
Dockerfile fix(docker): bump nginx runtime to 1.31.5-alpine3.24 and pin digest to resolve critical CVEs (#39561) 2026-09-03 08:46:07 -07:00
nginx.conf fix(ui): boot the UI image as an arbitrary uid by anchoring nginx writes under /tmp (#37982) 2026-08-24 11:57:36 -07:00