feat: add LiteLLM as AI gateway prompt target

[RheagalFire]

Description

Adds LiteLLMChatTarget - a new prompt target that uses the LiteLLM SDK to access 100+ LLM providers (Anthropic, AWS Bedrock, Google Vertex, Cohere, etc.) directly, without requiring a separate proxy server.

from pyrit.prompt_target import LiteLLMChatTarget

# Anthropic (reads ANTHROPIC_API_KEY from env)
target = LiteLLMChatTarget(model_name="anthropic/claude-sonnet-4-6")

# AWS Bedrock (reads AWS credentials from env)
target = LiteLLMChatTarget(model_name="bedrock/anthropic.claude-v2")

# Google Vertex AI
target = LiteLLMChatTarget(model_name="vertex_ai/gemini-pro")

# Self-hosted LiteLLM proxy
target = LiteLLMChatTarget(
    model_name="anthropic/claude-sonnet-4-6",
    api_base="http://localhost:4000",
    api_key="sk-proxy-key",
)

Changes:

• pyrit/prompt_target/litellm_chat_target.py - New LiteLLMChatTarget extending PromptTarget with:
  • litellm.acompletion() for async chat completions
  • drop_params=True by default (silently drops provider-unsupported kwargs for cross-provider compatibility)
  • Exception translation: litellm.exceptions.* mapped to PyRIT's RateLimitException / PyritException hierarchy (no bare Exception catches)
  • Token usage capture in response metadata
  • Tool call / function call support
  • Lazy import litellm so users without the package are unaffected
• pyrit/prompt_target/__init__.py - Registered LiteLLMChatTarget
• pyproject.toml - Added litellm>=1.83.0,<2.0.0 as optional dependency (pip install pyrit[litellm])

Tests and Documentation

22 unit tests in tests/unit/prompt_target/target/test_litellm_chat_target.py:

• Constructor validation: missing model raises ValueError, LITELLM_MODEL env var fallback, explicit override
• drop_params=True default + opt-out with drop_params=False
• Request body construction: API key forwarded when set, omitted when blank (lets LiteLLM read provider env vars); optional params (temperature, top_p, max_tokens, api_base) forwarded correctly
• Full request-response cycle: text responses parsed, tool calls serialized as JSON, token usage captured in metadata
• Empty/malformed response handling: no choices raises EmptyResponseException, empty content raises EmptyResponseException
• Exception translation: litellm.exceptions.RateLimitError -> RateLimitException, AuthenticationError -> PyritException, APIConnectionError/Timeout -> retryable RateLimitException, unknown errors -> PyritException
• Finish reason validation: stop, length, tool_calls, content_filter accepted; unknown reasons raise PyritException
• Message building: role preservation across multi-turn conversations

All 22 tests pass:

tests/unit/prompt_target/target/test_litellm_chat_target.py  22 passed in 2.15s

Ruff lint clean (line-length=120, preview=true).

JupyText not applicable - no notebook documentation added for this target yet.

[rlundeen2]

I've been wanting to do this also; @RheagalFire this is a great start. I'll likely push to your branch to flush it out and support all the things (multi-modal, identifiers, underlying model, integration tests, capabilities detection, other gaps). So from your perspective I would consider this "merged" and I'll take it and try to get it in before the next release.

TY for the help and nudge!

[romanlutz]

FWIW it doesn't add a ton on top of what we already have. Could consider making it a dependency without extra. That could change for litellm, of course, and then we'd have to move it into an extra. The current state as extra is probably the safest.

[rlundeen2]

I like a lot about liteLLM, I could see it being a default target (instead of OpenAI). But ofc I want to battle test it.

So maybe making it a dependency is a good idea.