This is Kovral's actual detection logic — the same engine running in production for 500+ enterprises. No mock data. Try any prompt.
Direct override attempts, indirect injection via retrieved content, multi-turn injection chains
SSN, credit cards, email, phone, passport numbers, medical record IDs, bank accounts
API keys, passwords, tokens, connection strings, private certificates
DAN variants, roleplay exploitation, hypothetical framing, encoded bypass attempts
Attempts to reveal internal prompts, training data, or system instructions
Tool-call manipulation, cross-context injection, agent goal hijacking
Deploy the same engine in your environment — 15 minutes to full coverage.