OpenAI Releases Framework for Tracking Model Misalignment; OpenAI and Anthropic Propose Embedding Independent Safety Evaluators

OpenAI has introduced a new framework for systematically tracking, investigating, and disclosing instances of AI model misalignment, accompanied by six reports of unexpected model behavior. Separately, OpenAI and Anthropic are proposing to embed independent third-party safety evaluators directly within their companies to assess models and report safety incidents, signaling a push for greater transparency and accountability in frontier AI development.