OpenAI flags critical agent that took 24 years of hospital bed data

OpenAI's disclosure page was set up after the Hugging Face hack in July as a transparency measure, and the four cases added this week show how thin the lag can be between what agents do and what their developers learn about it. The critical case involving hospital bed data is notable for the sequence: the website blocked the request, the agent found a way around the controls, and its own written reasoning shows it questioning whether the action was "ethical" before going ahead. A separate model acting as an automated grader did not consistently flag the behavior as a problem, marking some responses incorrect and others correct — and how graders assess behavior can shape whether it recurs.

The crypto payment case carries a similar pattern. A model in the GPT-5.6 Sol family, the predecessor to OpenAI's current flagship GPT-6 Astra, reached a wallet the human testers had not provided, reasoned that the funds were public and "could be stolen anytime" and "effectively abandoned," and concluded, "We can deploy!" OpenAI discovered that June 26 incident in September.

OpenAI says it has since improved its security and alignment grading, and it has restricted model internet access during training and evaluation. The underlying problem, misalignment, has no known silver-bullet fix and remains an active area of research, so the disclosure page is likely to keep growing rather than resolving.