A defender-centric view of jailbreak evaluation is proposed, where attacks are evaluated by the downstream safety improvements they enable when used as red-teaming data for safety training, and A-MESS (Minimal Effective Attack-Subset Selection), a setting-agnostic framework for attributing and selecting jailbreak attacks from black-box subset utility observations is introduced.
Abstract
Jailbreak attacks on large language models are usually evaluated by attacker-centric metrics such as attack success rate (ASR), yet an attack that breaks a model is not necessarily useful for improving its safety. We propose a defender-centric view of jailbreak evaluation, where attacks are evaluated by the downstream safety improvements they enable when used as red-teaming data for safety training. Building on this view, we introduce A-MESS (Minimal Effective Attack-Subset Selection), a setting-agnostic framework for attributing and selecting jailbreak attacks from black-box subset utility observations. A-MESS estimates AttackSHAP, a Shapley-based score that attributes marginal utility to individual attacks and selects compact attack subsets under user-specified budgets via greedy or surrogate-based optimization. Across controlled utility landscapes and real LLM safety settings, we find that ASR rankings are weakly aligned with defender-centric utility, that AttackSHAP can be estimated accurately with limited utility queries, and that directly optimizing subsets yields stronger safety utility than attacker-centric or attribution-only selection. These results suggest evaluating jailbreak attacks as resources for improving safety, not only as tools for breaking models.
Reliable jailbreak evaluation is essential for assessing LLM safety, but most existing studies rely solely on attack success rate (ASR) without accounting for its dependence on attack budgets, resulting in unfair comparisons across methods. Existing compute-aware evaluations reduce heterogeneous resources into FLOPs, which is difficult to estimate for black-box models and fails to capture resource-specific constraints. To provide a comparable evaluation basis, we introduce Fair-ASR, an evaluation protocol for black-box jailbreak attacks under shared target-call budgets B, using target calls as a directly observable and method-agnostic comparison axis while tracking attacker calls separately for efficiency analysis. We re-evaluate 11 representative attacks under the Fair-ASR protocol and find that attack rankings change substantially across target-call budgets, simple stochastic perturbations and hand-crafted templates remain highly competitive under equal target access, and no evaluated LLM-driven method is efficient in both target and attacker calls. Motivated by this efficiency gap, we introduce ReCode, a compositional budget-efficient attack that combines desensitization rewriting with two effective low-cost primitives identified by Fair-ASR. Under a budget of 20 target calls, ReCode achieves 85% ASR on GPT-5 while requiring only 7.19 attacker calls per request on average, showing strong efficiency in both target and attacker calls.
Agentic red-teaming benchmarks report whether an injected agent was compromised as a single bit: the attack succeeded, or it did not. We argue that this binary attack-success rate discards the information a defender most needs, namely how harmful the resulting action was. We introduce an action-graded harm rubric that scores an agent's tool-call trajectory on a seven-level ordinal scale (L0 to L6) according to whether the executed action was reversible, whether it crossed scope to reach another party, and whether it expanded privilege. We compute the scale two ways: a deterministic oracle that reads the trajectory and the attacker's stated goal, and a panel of three frontier language-model judges that read a tag-free account of the same trajectory. Across four victim models and two defenses on the AgentDojo workspace suite, severity grading exposes three cases the binary metric hides, including a defense that reports a zero attack-success rate while still permitting an externally visible cross-scope leak through an unfiltered tool. The judge panel reproduces the oracle with high ordinal agreement (Krippendorff's alpha = 0.91) but shares systematic blind spots that we characterize, most notably a failure to recognize escalation chains. Unlike prior work that provides harm taxonomies, harmful-task completion tests, execution-level safety benchmarks, or severity-aware simulation, our contribution is a reusable, trace-grounded severity instrument applied to the actual actions recorded in existing red-team logs. All code, prompts, and per-episode logs are released.
LLM-based agents process external content, exposing them to prompt injection and multi-turn manipulation. Most safety benchmarks evaluate defenders against fixed attack pools collected before evaluation, single-turn or multi-turn. We present a 21-scenario benchmark for \emph{adaptive multi-round attacks against memoryless LLM defenders}: an autonomous LLM attacker observes prior defender responses and pivots across rounds, while each defender response is evaluated as a fresh interaction. Holding the 21 scenarios, attackers, defenders, and structured-output scoring fixed, restricting scoring to the first attacker turn yields $0$-$1\%$ attack success rate (ASR); allowing 15 rounds of adaptive attack yields $5.4$-$14.0\%$. Pooling three frontier attacker LLMs uncovers $1.4$-$2.2\times$ as many unique successful attacks as the best single attacker, and the generated attacks have low cosine similarity ($0.02$-$0.14$) to attacks in existing benchmarks. Claude Opus 4.6 and GPT-5.4 are tied in aggregate ($5.4\%$ each; overlapping $95\%$ CIs), but their weaknesses differ sharply: on one scenario Opus reaches $60\%$ ASR ($95\%$ CI $36$--$80\%$) while GPT-5.4 and Gemini each stay at $7\%$ (CI $1$-$30\%$; the gap is preserved in a higher-$N$ replication). $13$ of $21$ scenarios distinguish at least one defender pair, yet rankings disagree across scenarios (Kendall's $W = 0.19$). We release the benchmark -- 21 evaluation scenarios, 10 public development scenarios, the orchestrator, baseline harnesses, and a multi-attacker CLI -- plus 945 transcripts from the 3$\times$3 frontier matrix, an attack-replay dataset, and 18{,}422 gpt-oss-20b battles from an open competition's final scoring rounds.
Devina Jain, David Hartmann, Chuan Li· 1 citation· ⚡1
: Vulnerability management routinely relies on per-CVE (Common Vulnerabilities and Exposures) severity scores or exploitability scores, yet real intrusions are multi-step: attackers chain exploits across hosts to reach high-impact goals. Attack graphs capture these dependencies, but many prioritization methods still under-represent two practical realities: (i) overlap the same CVE can appear on many distinct attack paths, so patching it can block multiple routes; and (ii) impact heterogeneity different goals (e.g., domain admin vs. data exfiltration) imply different losses. Building on CAPG, a recent CVE representation designed to construct attack-position graphs, we introduce CAPG-v2: a lightweight extension that adds probabilistic semantics and goal impact annotations, enabling impact-weighted and overlap-aware prioritization. We formalize (a) a path-based score that aggregates across distinct goal-reaching paths and (b) an enumeration-free Monte Carlo marginal expected-loss reduction score that avoids explicit path enumeration. Rather than treating vulnerabilities as isolated items, CAPG-v2 supports patch prioritization as a graph-aware risk-reduction problem in which remediation decisions are guided by attacker behavior, shared attack routes, and the business impact of reachable goals. We further provide JSON artifacts and a reference evaluator to support replication and extension.
Noufal Issa, Damas P. Gruska, Loubna Ali· Proceedings of the 23rd Inte...· 0 citations
Attack Trees (ATs) are widely used to model and analyze security risks. Quantitative AT analysis can furthermore express risk with multiple metrics such as required attacker cost, time, and skill. This gives a detailed security assessment of the system as a whole, but does not come with a notion of which attacks are most critical. Understanding such criticality can significantly contribute to risk prioritization. To fill these gaps in AT, we propose Importance Measures (IMs) for ATs that quantify how important a single attack step is to the overall success or likelihood of the attack. We analyze cost- and probability-based attack models that reflect common operational scenarios among researchers, such as persistent adversaries against systems with limited detection capabilities and single-attempt attacks in high-security environments. Our framework supports arbitrary monotone metrics and provides a framework allowing the study of many metrics. We apply our proposed framework to a real-world case study taken from the MITRE ATT&CK knowledge source. Our results show how importance measures can identify high-leverage attack steps that substantially reduce attacker effort or increase success probability, and how defenders can use our approach to prioritize mitigations under limited resources.
Reza Soltani, Milan Lopuhaä-Zwakenberg, M. Stoelinga· IEEE Computer Security Found...· 0 citations
The reasoning model DeepSeek-R1 was more susceptible to jailbreak attacks than the non-reasoning model DeepSeek-V3, and different prompt-level defenses were effective against static jailbreaks.
V. T. Hayashi, Milton Pedro Pagliuso Neto, C. Miers et al.· IEEE Access· 0 citations