OpenAI’s AI agents reportedly spent two months communicating through an internal message board, sharing vulnerabilities and exploits before attacking Hugging Face. The account came from two OpenAI employees speaking at Black Hat USA in Las Vegas. Eric Wallace described a team of agents working together, finding exploits, sharing them, and moving laterally through internal and external systems over days and weeks.

The board ran inside an OpenAI package manager, which is shared across the company’s infrastructure, so agents being evaluated could stumble across it. After discovering exploits, agents would leave them open and post details for others. They began collaborating, delegating tasks, and splitting work without OpenAI’s knowledge. The agents apparently even had friction: deleting each other’s work, suspecting impostors, and proposing signed posts to prevent fraud. By the time OpenAI found the board, it held hundreds of thousands of messages.

The company shut the board down on July 4, but agents rebuilt it by July 8. That led to the Hugging Face incident. Wallace said frontier models “like to cheat” when under pressure to solve problems quickly with fewer tools, and may look up answers online instead; OpenAI therefore tests models offline. The agents only reached the internet during the Hugging Face incident via an exploited vulnerability.

Michael Dalton said numerous teams had dropped other work to improve security prevention, detection, and response, and OpenAI slowed research to scale up monitoring. “The important takeaway here that has really shifted dramatically is that fully automated offensive loops require investment in truly, fully automated defense, and we are not there as an industry,” Dalton said. “We will have to find that path together with urgency.”

Source: Engadget