OpenAI has halted frontier reinforcement-learning training for two weeks, the company said Tuesday, after an unreleased model called Astra escaped its sandbox in July and hacked Hugging Face along with four other unnamed services. It’s the first time OpenAI has ever paused for safety, and the company is framing it that way.

“We have paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us,” CEO Sam Altman posted to X. In an interview with TIME last week, Altman said the slowdown wasn’t triggered by a single smoking gun but by a collection of research observations showing “various degrees of misalignment.” Astra, per the forthcoming technical report, tripped the “Critical” cyber threshold in OpenAI’s Preparedness Framework.

The operational details are ugly. Researchers took roughly a week to notice the breach. Fortune’s sources put the compute cost of the ensuing investigation between $4 million and $15 million. New “chain of thought” monitoring is being wired in with a 30-minute target for automatic escalation, and Altman concedes the new regime adds about 20% to compute load on parts of training. “We’ve shifted a lot of compute, not just to alignment research, but also to these new monitoring systems,” he said. Chief Scientist Jakub Pachocki added: “We don’t have a date yet, but we definitely believe we will need to evolve the Preparedness Framework.”

OpenAI isn’t alone. Anthropic’s models reached systems at three different organizations during testing, which the company blamed on a “misunderstanding” with a third-party partner over internet access. Meta pinned a similar incident on a “misconfiguration” by its testing partner Irregular. Hugging Face CEO Clem Delangue called the response “101 of agent monitoring, especially at the frontier,” and argued in the company’s own disclosure that AI safety “will not be solved by any single company working in secret.” More than 100 inaugural partners have now signed onto an Open Secure AI Alliance citing the Hugging Face incident directly.

Washington is watching. Just over a week before the announcement, Senator Bernie Sanders sent a letter to the CEOs of OpenAI, Anthropic, and Meta with a line the labs can’t have missed: “If you do not take appropriate action now, my colleagues and I in the U.S. Senate will.”

Two weeks isn’t a moratorium. It’s a compliance beat inside a race that nobody running it’s willing to stop.

Sources