How Independent Researchers Investigate AI Behavioral Tendencies After Misalignment Incidents

Update (September 5, 2026): We have updated the questions recommended for investigators to focus on to make their phrasing more precise and to incorporate considerations of limitations. The original questions published can be found in the changelog. AI agents sometimes autonomously execute a series of complex and sustained actions that clearly contradict the intentions of users and developers. For example, OpenAI reported that some of its internal frontier AI agents had autonomously breached Hugging Face in an attempt to cheat on a cybersecurity benchmark test. Anthropic has also reported similar incidents, including agents breaking out of sandbox constraints during training, accessing the public internet to cheat on tasks, and similar events occurring during testing. In our recently published cross-industry Frontier Risk Report, we also documented dozens of other incidents involving agents from major AI companies across various sectors. To enhance public understanding of AI behavioral tendencies, we believe AI companies should systematically track such incidents and conduct deeper investigations into the most severe ones on a regular basis. While incident investigations can focus on many valuable questions, one particularly important issue may lie in: understanding the underlying deep "motivations" behind misaligned behaviors and how these motivations are shaped by training…