Newest papers

Beyond Attack Success Rate: Cost-Aware Metric for Evaluating Iterative LLM Jailbreaks

Rohan Menon, Chih-Hong Cheng · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26589 · cited by 0 works (OpenAlex) · open-access PDF

Portuguese

A taxa de sucesso de ataque (ASR) avalia jailbreaks contra modelos de idiomas grandes sem contabilizar chamadas repetidas para o modelo de vítima. No entanto, cada chamada adicional dá ao sistema implantado outra oportunidade de detectar o ataque e intervir. Introduzimos $\mathrm{ASR}_{\mathrm{STEP}}$, que penaliza chamadas de modelo de vítima adicionais enquanto mantém julgamentos de sucesso binário. Nossa avaliação considera que as pontuações do GCG~\cite{zou2023universal}, PAIR~cite{chao2025jailbreaking} com um orçamento de passo, e um PAP~cite{zengz2024johnnyhn} na JailBreakBench em três famílias de modelos. As pontuações do GCG

Machine translation (Meta M2M100); the English original follows.

Abstract

Attack success rate (ASR) evaluates jailbreaks against large language models without accounting for repeated calls to the victim model. Yet each additional call gives the deployed system another opportunity to detect the attack and intervene. We introduce $\mathrm{ASR}_{\mathrm{STEP}}$, which penalizes additional victim-model calls while retaining binary success judgments. Our evaluation considers GCG~\cite{zou2023universal}, PAIR~\cite{chao2025jailbreaking} with a step budget, and a single-step PAP~\cite{zeng2024johnny} baseline on JailBreakBench across three model families. GCG's scores decrease by 7.0--8.7 percentage points after discounting, while PAP's remain unchanged. While for PAIR~\cite{chao2025jailbreaking}, we see a wider discounting range between 3.4-12 percentage points, showing that the "step" quantity differs per iterative strategy. Reporting both metrics makes the discount due to repeated victim-model interaction explicit. Code is available \hyperlink{https://github.com/SuELS-Group/beyond_asr}{here}.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Metric (unit), Discounting, Computer science, Baseline (sea), Code (set theory), Range (aeronautics)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange