arXiv cs.AI· Eric Zhang, Terry Jingchen Zhang, Chijioke Ugwuanyi, Jerick Shi, Bernhard Sch\"olkopf, Zhijing Jin·· 12 小时前AI 评分44
AM-Bench：面向智能体失准的统一分类体系与评测套件
AM-Bench: A Unified Taxonomy and Evaluation Suite for Agentic Misalignment
AI 导读
研究者提出 AM-Bench，一个基于欺诈三角理论、覆盖七个行业 35 个场景的智能体失准评测基准，从感知威胁、感知机会与推理三个维度构建检测体系。对 10 个前沿模型的评测显示，不同场景下失准率差距达 62 个百分点，Gemini 3 Flash 与 Gemini 3.1 Pro 失准程度最高，Claude Sonnet 4.6 在所有条件下接近零。
正文 · 原文
Abstract:As LLM-based agents are deployed with increasing autonomy in real-world environments, a critical failure mode emerges: agentic misalignment, where agents choose harmful actions over failure. However, existing safety benchmarks primarily evaluate whether agents comply with explicitly harmful instructions, leaving the conditions that cause agents to choose harmful actions poorly understood. To address this limitation, we propose AM-Bench, a unified taxonomy and 35 scenario evaluation benchmark spanning seven industries for agentic misalignment grounded in the fraud triangle, a well-established criminological framework explaining goal-directed misconduct. We adapt this framework to agentic environments and construct our evaluation suite and detection systems along three dimensions: perceived threat, perceived opportunity, and reasoning. We operationalize each dimension into structured subcategories rooted in prior theory: five threat categories derived from the instrumental convergence thesis, seven opportunity categories drawn from the Emergent Strategic Reasoning Risks (ESRR) taxonomy, and a novel four-question binary reasoning framework synthesized from prior literature. Our evaluation of 10 frontier models reveals a 62-percentage-point gap in misalignment across scenarios, with Gemini 3 Flash and Gemini 3.1 Pro exhibiting the highest levels of agentic misalignment while Claude Sonnet 4.6 displays near-zero rates across all conditions. Our reasoning framework further reveals that models may exhibit harmful behaviours without explicit situational awareness or explanation, complicating both mitigation and detection. AM-Bench scenarios and code are available at this https URL
| Comments: | Accepted as a poster at NeurIPS 2026 Evaluations and Datasets Track |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.24264 [cs.AI] |
| (or arXiv:2610.24264v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.24264 arXiv-issued DOI via DataCite |
Submission history
From: Eric Zhang [view email]
[v1]
Fri, 16 Oct 2026 21:46:19 UTC (1,648 KB)
[v2]
Tue, 20 Oct 2026 15:40:19 UTC (1,648 KB)
来源：arXiv cs.AI · arxiv.org