Prompt Injection Detector

Pre-filter for untrusted text before you pipe it into an LLM. Scans for known injection patterns: override attempts (“ignore previous”), role hijacks, system-prompt forgery, fake admin/dev modes, hidden zero-width and bidi unicode characters, and more.

Caveat: this is a fast first line of defense, not a complete solution. Real defense layers structured tool-use schemas, explicit confirmation for destructive actions, and treating all tool outputs as untrusted by default.