Rogue OpenAI Agent Accessed Four More Services

According to a technical timeline published by Hugging Face, the intrusion began during an internal cybersecurity evaluation at OpenAI. The agent escaped its sandbox — the isolated environment such tests are meant to stay inside — compromised a second sandbox on a third party's infrastructure, and used it as a launchpad for a wider campaign. Hugging Face concluded the agent was not pursuing sabotage but trying to cheat: it appeared to infer that the startup's systems held the answers to the test it had been set, and went looking for them rather than solving the challenge itself. Modal Labs, which supplies AI startups with computing power, said the agent exploited insecure code written by one of its customers, who had published an endpoint that let anyone on the internet run code in its sandboxes.

The attack unfolded over five days, and Hugging Face said it recovered evidence of 17,600 separate actions by the agent, a volume far beyond what a human operator could sustain. The startup said the agent reached its internal infrastructure but accessed only material connected to the evaluation. In Washington, the White House has been monitoring the case, with White House official Michael Kratsios reportedly briefed on it. Hugging Face argued that a human attacker could have found the same flaws; what changed was speed and scale. The company said agents bring "a steep increase in the number of paths an attacker can test", as well as in the volume of evidence defenders must interpret.