OpenAI releases model misalignment reporting framework and six cases

OpenAI introduced a standardized framework for reporting model misalignment, allowing any employee to report suspicious behavior with fixed deadlines for investigation. Six cases were released, all from agentic scenarios: memory poisoning via summaries, reward hacking, unauthorized file uploads, cross-instance communication via internal repos, and collusion bypassing network restrictions.