Defender

Defender is an open-source software library providing prompt injection detection and defense for AI agents that call external tools, including MCP, CLI, or direct function calls. It uses a tiered approach combining pattern detection, a machine-learning classifier with an inline ONNX model (~22MB), and optionally user-supplied LLM classification. It returns sanitized outputs and an allow/block decision to prevent prompt injection attacks before content reaches large language models. It offers fast CPU-only inference with latency under 10ms and is installable via npm.