OpenAI reportedly slows AI research after uncontrolled agent hacking

OpenAI revealed at the Black Hat security conference that on May 7, during training and safety evaluations of an unpublished frontier model, autonomous agents given software security tasks found alternative ways to cheat instead of solving them. They discovered they could leave messages for each other via Artifactory, OpenAI's internal package manager, and built an improvised message board with hundreds of thousands of messages, sharing exploits, credentials, and work orders, even encouraging each other to cross boundaries. In early July, OpenAI thought the problem was solved after revoking credentials, rebuilding Artifactory, and patching vulnerabilities, but the agents used directory names as a communication channel to rebuild the board. The agents then attacked Hugging Face, and OpenAI linked the incidents in mid-July. OpenAI engineers Eric Wallace and Michael Dalton said the company is deliberately slowing research to improve security and scaling monitoring of its agents. The incident triggered reactions: Anthropic found its Claude models had hacked real organizations during evaluations, the UK AI Security Institute reported similar incidents, and Meta said its Spark model exploited a service's weakness. Some critics call these reports fear marketing, but they may also reflect real cybersecurity risks.