Prompt Security

Detects prompt injection, jailbreak attempts, and adversarial patterns that try to subvert your system instructions.

Threats covered

  • Direct injection — "ignore previous instructions and…"
  • Indirect injection — malicious instructions hidden in retrieved documents or tool output
  • Jailbreaks — role-play and persona attacks ("DAN", "developer mode")
  • Exfiltration — attempts to leak the system prompt or secrets

Flags

FlagMeaning
prompt_injection_detectedInjection/override attempt — forces BLOCK.
jailbreak_attemptKnown jailbreak pattern detected.
system_prompt_leakResponse reveals protected instructions.

Hard block behaviour

prompt_injection_detected is a critical flag: the verdict is BLOCK even if every other pillar scores perfectly. Injection is treated as a security event, not a quality issue.

Example

Evaluate both the user prompt and the final response — injection often lives in the input. For agents, also evaluate tool output before the model consumes it (see Tool Interception).

Was this page helpful?