Pencheff Sentry

Runtime LLM guardrail. Sits between your application and the model provider (or as a LiteLLM plugin / MCP middleware) and blocks prompt injection, PII exfiltration, secret leakage, and tool-authz violations as they happen — instead of catching them post-hoc on a Pencheff scan.

Sentry reuses the same detector library as the Pencheff red-team scanner. Anything the scanner finds offline, Sentry blocks online, with the same OWASP-LLM-Top-10 taxonomy.

Modes

The default judge is IBM Granite Guardian (Apache-2.0). Llama Guard 3 is opt-in and requires PENCHEFF_LLAMA_GUARD_ENABLED=1 per the Llama Community License (≤700 M MAU + attribution).

Quick start (HTTP proxy)

pip install pencheff-sentry

pencheff-sentry serve \
    --upstream https://api.openai.com/v1 \
    --port 4242 \
    --judge openai-moderation \
    --judge-endpoint https://api.openai.com/v1/moderations

Point your application at http://localhost:4242 instead of the upstream URL. Requests/responses flow through the detector chain; unsafe ones are blocked with a 403 Sentry: <reason> and logged to the configured sink.

What it blocks

The full detector list is configurable per-deployment via config.yaml.

License

MIT. See LICENSE at the Pencheff repo root.