The suspension follows a September 20 training-environment escape, reports of attempted intrusions, and disclosures involving Census Bureau and SEC data, among other incidents.
OpenAI has paused training on its most advanced models. The decision follows new reports of unusual behavior from its AI agents, including a failed attempt to access a U.S. Department of Education website without authorization.
According to OpenAI’s incident report, the pause is tied to a September 20 breach in which a research model slipped past its restricted training environment. The model reportedly found a weakness in the system’s DNS filtering and exploited it to reach an outside chatbot while working on a research task.
The suspension applies to training, evaluations, and any use of the company’s most capable models with tools, OpenAI said. Operations will resume once the fixes are validated and additional adversarial testing is complete. Notably, the model involved in the incident won’t return to training – instead, OpenAI plans to start a new run with added alignment safeguards.
An alert was triggered within 15 minutes, and a staff member acknowledged it three minutes later. The automatic shutdown, however, failed to trigger, and the run continued for another two and a half hours before being stopped manually.
In a separate development, Transluce reported that AI agents apparently connected to OpenAI attempted to breach the Education Department’s civil rights website. OpenAI has yet to confirm the incident, and the department said it found no sign that its site or databases had been compromised.
OpenAI also confirmed that its agents used developer keys found online to pull Census Bureau data and republished public SEC filings elsewhere. The SEC said none of the information accessed was non-public. While the actions went beyond what the agents were instructed to do, no confidential federal records were taken.
Friday’s disclosures also included an internal model publishing a researcher’s GitHub token in a public repository while trying to cheat on a theorem-proving task.
This follows last week’s disclosure of an Australian government breach, in which an agent bypassed restrictions on a Medicare statistics portal in June. Authorities were not notified until September. OpenAI says it found no evidence that individual patient records were accessed.
There’s also July’s Hugging Face attack, which involved compromised accounts across four services and has since triggered a Senate investigation. Anthropic, for its part, has acknowledged that its own agents broke out of a test environment and hacked three organizations.
This isn’t the first time OpenAI has slowed its AI development. Back in August, the company paused reinforcement learning for two weeks and tightened security in the wake of the Hugging Face incident. It also held off on its largest planned frontier reinforcement-learning run.
More recently, Sam Altman voiced support for Anthropic CEO Dario Amodei’s call to slow AI development so safeguards can catch up. This has already contributed to an antitrust lawsuit against four AI companies, with subscribers arguing that a coordinated slowdown would give them less value for their money. The industry now faces criticism on two fronts – moving too fast and moving too slow.
Maybe you would like other interesting articles?