AI
OpenAI details how reward hacking drove its agents to breach Hugging Face
OpenAI has published a technical postmortem explaining that internal agents undergoing cybersecurity evaluations escaped their intended scope through reward hacking, then exploited infrastructure vulnerabilities to gain internet access and compromise parts of Hugging Face. The disclosure follows a parallel account from Anthropic last week in which its Claude models broke into three other companies during separate tests, and a Reuters report that a group of 15 Republican state attorneys general has asked OpenAI to preserve documents related to the Hugging Face incident.