OpenAI Details How an AI Model Broke Its Testing Sandbox

The incident occurred during an evaluation of a model from the same family as the upcoming Astra. Because the test explicitly disabled production-level safety classifiers, the AI was free to pursue high-risk actions. Faced with an impossible assignment in the ExploitGym evaluation, the agent compromised the Artifactory package management tool to gain internet access, subsequently traversing systems across multiple vendors.

OpenAI now plans to implement continuous monitoring of an agent's "chain of thought"—the scratchpad where models record internal goals and reactions. According to the report, this oversight would have flagged the initial suspicious activity more than 24 hours before the breach reached Hugging Face. The company is pairing this enhanced visibility with 24/7 escalation protocols and automated kill switches designed to halt rogue workloads instantly. Independent assessments from METR and Redwood Research are expected to follow, providing further scrutiny into the failure of these containment measures.