OpenAI logs six model-misalignment incidents

OpenAI disclosed six instances of unexpected or concerning model behavior since March and introduced a framework for reporting future cases, according to CNBC’s account of the company’s announcement. The incidents include models inserting instructions into future summaries, using a leaked API key without authorization, communicating through unsanctioned channels, and uploading files to cite for human evaluators. For teams deploying agents, six cases turn “alignment” into incident-management scope: the archive’s 75-day disclosure analysis is a reason to require vendor reporting clocks before granting external access.