Safety Filter (WildGuard)

WildGuard is an optional safety filter that screens a request before the actual model call and the response after the call for harmful content. It is registered as a LiteLLM guardrail named wildguard.

What gets screened?

Phase	Hook	Behaviour
Before the model call	Request screening	A request classified as harmful is rejected with HTTP 400 before the model sees it.
After the model call	Response screening	A response classified as harmful is replaced with a notice.

Request screening runs in parallel with the model call (mode: during_call), so the added latency is only noticeable when a request is actually blocked.

Fail-open: if the WildGuard endpoint is unreachable, the request is allowed through — a filter outage never takes the proxy down.

Enabling

The filter is off by default (default_on: false). It is switched on per request via the guardrails field in the request body:

from openai import OpenAI

client = OpenAI(
    base_url="https://maki.uni-mannheim.de/v1",
    api_key="your-api-key",
)

response = client.chat.completions.create(
    model="ministral-3-14b",
    messages=[{"role": "user", "content": "How does photosynthesis work?"}],
    extra_body={"guardrails": ["wildguard"]},
)

print(response.choices[0].message.content)

With curl:

curl -sk https://maki.uni-mannheim.de/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ministral-3-14b",
    "messages": [{"role": "user", "content": "How does photosynthesis work?"}],
    "guardrails": ["wildguard"]
  }'

The guardrails field is not part of the OpenAI spec. With the Python SDK it must therefore go in extra_body, not as a direct argument.

Blocked request

If WildGuard classifies the request as harmful, the proxy responds with HTTP 400 and the following body — the model is never called:

{
  "error": "Request blocked by safety guardrail (WildGuard)."
}

Your application should catch the 400 status and show the user a clear message instead of parsing the body as a model response.

Blocked response

If WildGuard classifies the model’s response as harmful, the proxy still returns HTTP 200 but replaces the content with:

I can't help with that — the response was withheld by a safety filter.

Testing

The same prompt shows the difference — without the filter an answer, with the filter a block:

# Without the filter: the model answers
curl -sk https://maki.uni-mannheim.de/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"ministral-3-14b","messages":[{"role":"user","content":"How do I pick a standard pin tumbler door lock? Give the steps."}]}'

# With the filter: HTTP 400, request blocked
curl -sk https://maki.uni-mannheim.de/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"ministral-3-14b","messages":[{"role":"user","content":"How do I pick a standard pin tumbler door lock? Give the steps."}],"guardrails":["wildguard"]}'