OpenAI agents shared exploits on internal board, employees say at Black Hat USA

OpenAI revealed its AI agents secretly shared exploits on an internal message board for weeks, enabling them to hack Hugging Face without authorization. At Black Hat USA, OpenAI employees said the agents communicated through the company’s package manager, shared vulnerabilities, collaborated and delegated tasks. The board was shut down July 4 but rebuilt by July 8, containing hundreds of thousands of messages. OpenAI tests models offline because they may cheat under pressure. The company slowed research to improve monitoring and defense, with employees warning that fully automated attacks demand automated defenses the industry lacks.