Shipped esys-watch
to PyPI — a local CLI that scans text for secrets, PII, and prompt injection patterns before it gets pasted into an LLM.
Why: as more of us route logs/code/docs through ChatGPT, Claude, Cursor etc, the leak surface is the copy-paste itself, not just the API layer. This runs entirely local, no account, nothing sent anywhere.
Bash
pip install esys-watch
esys-watch --demo
Validated against 320 labeled cases (100% recall, 0% FP), plus adversarial testing for ReDoS, malformed input, and concurrent file access.
Open source (MIT), still v0. If anyone here works on LLM security / prompt injection detection, I’d love feedback — especially on where the regex-based injection detection falls short (it’s intentionally scoped to classic patterns only, not multi-turn or indirect injection yet).