LLM defense SDK
Defense layer for LLM apps. 3-stage pipeline (regex, ML classifier, LLM judge) that catches prompt injection, data leakage, and jailbreaks in under 20ms. Span-level redaction, not binary blocking. Ships as Python SDK, self-hosted API, and MCP server.
Every LLM app I've worked on has the same vulnerability: untrusted input goes directly into prompts. The market validates the problem (Lakera $300M, Promptfoo acquired by OpenAI), but most solutions are cloud-only and expensive. The prompt injection at the top of this site is a demo of the problem.
Python SDK (pip install unplug), FastAPI server (unplug serve), MCP server. Auto-instruments LangChain, CrewAI, LlamaIndex.
v0.1 MVP (regex pipeline) is done. Next: ONNX classifier, framework integrations, SaaS launch.