A system-level response addresses security failures outside the prompt and model generation process. Use one when prompt hardening delays but cannot prevent persistent behavioral drift or gradual scope expansion. section 10.5.1
Possible responses include:
- Evaluate outputs turn by turn, enforce conversation scope, or filter content. section 10.5.1
- Audit conversations, route between models, or limit request rates. section 10.5.1
Persistent scope expansion
A conversation gradually shifts beyond its intended scope and elicits restricted content over several reasonable turns. Prompt hardening may delay this failure; consider conversation scope enforcement, turn-by-turn evaluation, or content filtering. section 10.5.1
