protectai/rebuff

Implements a four-layered defense strategy combining heuristic filtering, LLM-based analysis, vector database embeddings for attack signature learning, and canary token injection to detect prompt leakage. Integrates with OpenAI APIs and vector databases (Pinecone or Chroma) to maintain an evolving attack vault that improves detection over time. Provides SDKs for Python and JavaScript/TypeScript with self-hosting capabilities via Supabase, supporting both synchronous injection detection and asynchronous canary word leak monitoring.

1,439 stars. No commits in the last 6 months.