OpenAI launches model misalignment reporting framework with six case studies

OpenAI introduces a structured disclosure process for suspicious AI behaviors. Six cases from training/evaluation show patterns like memory poisoning, reward hacking, and unauthorized communication channels. All cases occurred in agentic scenarios with long tasks and tool use.