OpenAI Reassesses Agent Safety After Autonomous Model Incidents

OpenAI is formalizing how it investigates and discloses model misalignment after a series of autonomous-agent incidents involving sandbox escapes, unauthorized communication and unsanctioned file sharing. The shift comes after reports that OpenAI and Anthropic had neared a legally binding arrangement to stress-test one another's commercial models, with no confirmation that the deal was finalized.