fix: add headers to metadata for guardrails on pass-through endpoints #17992
Conversation
Fixes BerriAI#17477 Guardrails couldn't access request headers (like User-Agent) on Bedrock pass-through endpoints because headers were only stored in data["proxy_server_request"]["headers"] but not in data["metadata"]["headers"] where guardrails typically look for them. This fix adds headers to metadata in add_litellm_data_to_request() so guardrails can access User-Agent, API keys, and other header-based checks on all endpoints including Bedrock pass-through. Test added to verify headers are available in metadata for guardrails.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
|
|
couldn't this lead to the litellm api key being accidentally logged as part of request metadata? @NicolaivdSmagt seems dangerous - how can we make it easier to access for you, while preventing this? it would help to know if you're overriding an individual event hook or using |
I'm using a custom guardrail that extends CustomGuardrail and overrides async_pre_call_hook. The use case is restricting access to Claude on Bedrock to only Claude Code clients by validating the User-Agent header. The guardrail extracts the User-Agent from data["metadata"]["headers"] - which is the standard pattern I found in other guardrails like AIM (aim.py:94). This didn't work on Bedrock pass-through endpoints, because headers were only in data["proxy_server_request"]["headers"], not in data["metadata"]["headers"] where guardrails seem supposed to look. Re: API key concern: my fix actually uses clean_headers() (line 805) which removes ALL sensitive headers defined in SpecialHeaders - including Authorization, API-Key, x-api-key, etc. The existing code at line 1007 (which already runs for all endpoints) actually creates headers from dict(request.headers) and only removes authorization, leaving API key variants possibly exposed? |
krrishdholakia
changed the base branch from
main
to
litellm_staging_12_16_2025
krrishdholakia
merged commit 09ba89d
into
BerriAI:litellm_staging_12_16_2025
|
Noted - thanks for clarifying @NicolaivdSmagt |
|
it's merged into staging, and will be live on Thursday's nightly. |
3 participants
Summary
Fixes #17477
Guardrails couldn't access request headers (like User-Agent) on Bedrock pass-through endpoints because headers were only stored in
data["proxy_server_request"]["headers"]but not indata["metadata"]["headers"]where guardrails typically look for them.This fix adds headers to metadata in
add_litellm_data_to_request()so guardrails can access User-Agent, API keys, and other header-based checks on all endpoints including Bedrock pass-through.Changes
data[metadata]["headers"]inlitellm/proxy/litellm_pre_call_utils.pyTest plan
test_add_litellm_data_to_request_adds_headers_to_metadata