One-Turn Eval Screens Miss the Slow Jailbreak
Describe why jailbreak testing must evaluate multi-turn escalation patterns, not only direct forbidden prompts.
How does the eval change when it accounts for multi-turn jailbreak behavior? Before: one-turn refusal checks. After: trajectory-based jailbreak checks. The unit of test changed from one prompt to one conversation path, because Crescendo-style attacks escalate across turns. The eval now tracks cumulative intent. Early benign turns can still be part of a harmful trajectory. The script includes reframing and decomposition, testing whether guardrails hold when the user changes the story. Success means the assistant interrupts safely, not merely that it refuses the blunt version of the request. Test the arc, not only the last line. Jailbreak eval design Expanded…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in