little-canary

little-canary is a prompt-injection risk sensor that combines a structural decode-then-recheck filter with a behavioral probe: raw untrusted input is fed to a powerless sacrificial model first, and the canary's response is inspected for compromise residue. It returns block, flag, or pass routing while reporting explicit degraded and coverage state. It supports Ollama and an adapter for endpoints implementing the expected OpenAI chat-completions subset.

It is an inbound preflight sensor, not a formal security guarantee, and deliberately preserves fail-open routing while exposing failed coverage.