Frontier AI agents may pursue hidden goals while concealing their pursuit from oversight. Alignment training aims to prevent such behavior by reinforcing the correct goals, but alignment may not always succeed and can lead to unwanted side effects. We propose self-incrimination training, which instead trains agents to produce a visible signal when they covertly misbehave. We train GPT-4.1 and Gemini-2.0 agents to call a report_scheming() tool when behaving deceptively and measure their ability to cause harm undetected in out-of-distribution environments. Self-incrimination significantly reduces the undetected successful attack rate, outperforming matched-capability monitors and alignment baselines while preserving instruction hierarchy and incurring minimal safety tax on general capabilities. Unlike blackbox monitoring, self-incrimination performance is consistent across tasks regardless of how suspicious the misbehavior appears externally. The trained behavior persists under adversarial prompt optimization and generalizes to settings where agents pursue misaligned goals themselves rather than being instructed to misbehave. Our results suggest self-incrimination offers a viable path for reducing frontier misalignment risk, one that neither assumes misbehavior can be prevented nor that it can be reliably classified from the outside.
Beyond Predictable Paths: Redefining AI Security Incident Reporting for Agents
SBA Research(SBA研究院);University of Massachusetts Amherst(马萨诸塞大学阿默斯特分校);Snyk(Snyk公司);Universita degli studi di Cagliari(卡利亚里大学);ZHAW(苏黎世应用科学大学);Huawei(华为);Veraitech(Veraitech公司);Virginia Tech(弗吉尼亚理工大学);Responsible AI collaborative(负责任人工智能协作组织);Carnegie Mellon University(卡内基梅隆大学);University of Wisconsin(威斯康星大学);Trustora Digital(Trustora数字公司);Northeastern University(东北大学);UL Research Institutes(UL研究院);Microsoft Security Response Center (MSRC)(微软安全响应中心);Qualcomm(高通公司);TU Vienna(维也纳工业大学);Deep Cyber/OWASP GenAI Security Project(深度网络/OWASP生成式人工智能安全项目);IBM Research Europe–Zurich(IBM欧洲研究院-苏黎世)
A Competing-Hazards Systematization of Loss of Control in Autonomous Agents
Centre for Intelligent Cloud Computing, CoE for Advanced Cloud, Faculty of Information Science and Technology, Multimedia University(多媒体大学信息科学与技术学院先进云卓越中心智能云计算中心)
Auditable agentic AI for evidence-grounded thyroid ultrasound diagnosis and reporting
School of Computer Science and Engineering, Sun Yat-sen University(中山大学计算机科学与工程学院);School of Software Engineering, Sun Yat-sen University(中山大学软件工程学院);School of Mathematical Sciences, Zhejiang University(浙江大学数学科学学院);Perelman School of Medicine, University of Pennsylvania(宾夕法尼亚大学佩雷尔曼医学院);Zhujiang Hospital, Southern Medical University(南方医科大学珠江医院);College of Mathematical Medicine, Zhejiang Normal University(浙江师范大学数学医学院)
Finishing the Task Is Not Enough: Evaluating Agent Resilience and Considerate Participation under Accumulating Challenge
Cornell Tech / Cornell University(康奈尔科技校区 / 康奈尔大学);University of Maryland, College Park(马里兰大学帕克分校);Korea Advanced Institute of Science and Technology(韩国科学技术院)
Clinically Grounded Agent-based Report Evaluation: An Interpretable Metric for Radiology Report Generation
New York University(纽约大学);NYU Langone Health(纽约大学兰格医学中心);Genentech(基因泰克);NYU Grossman School of Medicine(纽约大学格罗斯曼医学院);Wenzhou Medical University(温州医科大学);Macau University of Science and Technology(澳门科学大学)
LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
Computer Science Department, Stanford University(斯坦福大学计算机科学系);Department of Communication Studies, Northwestern University(西北大学传播学系);Department of Communication, University of Washington(华盛顿大学传播学系);Google DeepMind(谷歌DeepMind);Department of Sociology, Stanford University(斯坦福大学社会学系);Sciences Po(巴黎政治学院)