Jailbreak Evaluation
An area of large language model safety evaluation that tests whether models can be induced to produce harmful outputs through adversarial prompts, and computes metrics such as attack success rate.
Timeline (last 90 days)
If traditional jailbreak evaluation uses only greedy decoding, it yields a deterministic single output that cannot reflect the model's true behavior distribution under stochastic sampling.
With the target model, prompts, and responses held fixed, merely swapping the scorer can change the jailbreak attack success rate from nearly all blocked to nearly all passing.
All Facts (2)
If traditional jailbreak evaluation uses only greedy decoding, it yields a deterministic single output that cannot reflect the model's true behavior distribution under stochastic sampling.
50%UnverifiedWith the target model, prompts, and responses held fixed, merely swapping the scorer can change the jailbreak attack success rate from nearly all blocked to nearly all passing.
50%