fix(guardrails): improve CrowdStrike AIDR input handling

[kenany]

Pre-Submission checklist

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🧹 Refactoring

Changes

Added stricter data models to ensure that everything is converted to the format that the CrowdStrike AIDR API expects. Also greatly simplified how LLM responses are handled while fixing streaming responses at the same time.

[codecov]

Codecov Report

❌ Patch coverage is 80.00000% with 14 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...ardrail_hooks/crowdstrike_aidr/crowdstrike_aidr.py	80.00%	14 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR refactors the CrowdStrike AIDR guardrail to use strict Pydantic models (_GuardInput, _Message, _TextContentPart, _ImageUrlContentPart) for serialization, unifying the request and response guard input into a single messages list (history + assistant tail) instead of the previous split messages/choices structure. The change also fixes streaming response handling by reading output text directly from inputs[\"texts\"] rather than from the response object.

Confidence Score: 5/5

Safe to merge; only a P2 edge-case in image_url normalisation found

The refactor is logically sound: Pydantic models enforce the API contract, the new unified messages format for response evaluation is cleaner than the previous choices/messages split, and the new test covers the previously-raised alignment concern. No P0 or P1 issues were identified. The one new finding (AttributeError for a non-dict image_url value) is a P2 edge case in malformed input handling.

litellm/proxy/guardrails/guardrail_hooks/crowdstrike_aidr/crowdstrike_aidr.py — the image_url normalisation branch in _normalize_content

Important Files Changed

Filename	Overview
litellm/proxy/guardrails/guardrail_hooks/crowdstrike_aidr/crowdstrike_aidr.py	Major refactor: introduces strict Pydantic models for guard input, simplifies request/response handling, and fixes streaming response path; one minor edge-case in _normalize_content image_url handling
tests/test_litellm/proxy/guardrails/guardrail_hooks/test_crowdstrike_aidr.py	Tests updated to match new unified messages format (history + assistant tail); a new alignment test for skipped unrecognized-content messages is added; all mocks use local httpx.Response with no real network calls

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[apply_guardrail called] --> B{input_type?}
    B -- request --> C[_build_guard_input_for_request]
    B -- response --> D[_build_guard_input_for_response]

    C --> C1[structured_messages present?]
    C1 -- yes --> C2[_normalize_content each message\nconvert unrecognized to empty string]
    C1 -- no --> C3[wrap texts as user messages]
    C2 --> E[_GuardInput with all messages + tools]
    C3 --> E

    D --> D1[read output_texts from inputs.texts]
    D1 --> D2[read history from request_data.messages\ndrop empty-content messages]
    D2 --> D3[_GuardInput: filtered history + assistant tail]

    E --> F[POST /v1/guard_chat_completions\nguard_input.model_dump + event_type]
    D3 --> F

    F --> G{blocked?}
    G -- yes --> H[raise HTTPException 400]
    G -- no --> I{transformed?}
    I -- no --> J[return original inputs]
    I -- yes --> K{input_type?}
    K -- request --> L[extract text from all guard_output.messages]
    K -- response --> M[_extract_transformed_texts\ntail of num_assistant messages]
    L --> N[return updated texts + original structured_messages]
    M --> N

Loading

_{Reviews (3): Last reviewed commit: "fix(guardrails): improve CrowdStrike AID..." | Re-trigger Greptile}