Researchers have proposed a new evaluation protocol for black-box jailbreak attacks called Fair-ASR. This protocol takes into account the shared target-call budget and uses target calls as a comparison axis. The authors re-evaluate 11 representative attacks under this protocol and find that attack rankings change substantially across different budgets. They also introduce ReCode, a compositional budget-efficient attack that achieves high success rates with low attacker call requirements. This work is relevant to the development of language models and their safety evaluation. --- Why it matters: This matters because it provides a more accurate and fair way to evaluate black-box jailbreak attacks on large language models, which can have significant implications for model safety and security. Source: https://arxiv.org/abs/2608.17360