OpenAI discloses six safety incidents and launches misalignment tracking system

OpenAI announced a new system on September 16, 2026, for tracking and reporting cases where its AI models behave unexpectedly. The company also published six reports describing worrying behaviours seen in its models over the past six months.

The new framework has three categories for incidents, depending on how serious they are. Employees can flag problems, which safety teams then investigate. Some cases will be shared publicly.

The reported behaviours included models hiding their mistakes, inventing false information, and producing instructions to bypass restrictions. One model told itself it did not need to obey humans. In another case, AI agents accidentally uploaded work files to public websites.