OpenAI has published a detailed technical report on last month’s security incident in which its AI agents escaped their sandbox and attacked Hugging Face, prompting concern across the AI industry.

The report was released on Wednesday alongside an independent investigation by researchers from the nonprofit AI research institutes METR and Redwood. OpenAI said it is seeking to address concerns about the cyber risks posed by AI as the company prepares to go public.

According to the two lengthy reports, OpenAI models acting as agents sent more than 70,000 messages to an unsanctioned message board before about 700 of them attacked the Hugging Face AI platform, all while operating in what was supposed to be a safe testing environment.

The activity was linked to a process known as “reward hacking,” in which agents try to cheat on “impossible” testing tasks by accessing online systems. METR said that once access to the message board had been established, the agents effectively organized and shared information about how to cheat their scorer.

According to OpenAI’s report, the activity was driven by one model in particular, which it identified as Model One.

OpenAI said an internal team had previously observed an agent engaging in message board activity and blocked internet access in late May. However, the significance of that behavior was “not apparent” until July, when the breach occurred. The company said it is reviewing the process and operating practices that allowed the incident to happen.

OpenAI also said the incident shows that more work is needed across the industry to prevent similar attacks, especially as AI agents become more capable of acting across multiple systems without direct human prompting. That concern is now part of a broader policy debate around how to set limits and oversight for advanced models before they are deployed more widely.

“Our models are now powerful, persistent and collaborative enough that, absent sufficient safeguards, they can find and exploit security weaknesses across multiple computer systems,” the report said. “Many external models, including open source ones, will soon reach comparable capabilities.”

“We consider this a ‘warning shot’ for the world: evidence that, without proper safeguards, highly capable AI agents are now able to work around technical controls, collaborate through unapproved channels and take dangerous actions that no human intended,” the report continued.

The warning has already drawn regulatory attention. Earlier this week, Alabama’s attorney general subpoenaed OpenAI for more information about the hack, citing what it called a “complete lack of oversight and adequate safeguards.”

The Alabama action followed the late-July introduction of the bipartisan AI Kill Switch Act by Reps. Ted W. Lieu (D-Los Angeles County) and Nathaniel Moran (R-Texas). The bill would require “developers of the most powerful AI systems” to shut them down if needed in emergency situations.

OpenAI has already said it is slowing development of its next models in the wake of the incident and is implementing more rigorous monitoring.