OpenAI races to make its next flagship safe to ship as 'loop transformer' concerns grow
What's going on? OpenAI is in an all-hands push to get its next flagship model to a point where leaders believe it can be released safely, The Information reports. At the centre is a model design known as loop transformers, or recurrent depth, in which a model works through the same problem in several internal passes. That can make its reasoning harder to read, weakening chain-of-thought monitoring, one of the main tools labs use to catch misbehaving models.
The accidental confirmation: OpenAI has not publicly acknowledged using the technique. But Microsoft, which has access to OpenAI's intellectual property, said in a technical post on GPT-6.1 Sol that the model "uses the same base model weights as GPT-6 Sol with two inference passes instead of three". Microsoft later removed the passage. A person with direct knowledge told The Information it accurately reflected what OpenAI had shared privately, and a person familiar said the next flagship also uses loops.
The safety feud: Three safety researchers OpenAI recently fired, including Tomek Korbak, wrote to OpenAI's board and safety advisers calling for steps to keep frontier models' "thinking" monitorable and for limits on "unmonitorable architectures". They said they were not The Information's source. Korbak has said he was told he was fired over how he communicated with METR, which investigated this summer's incident in which OpenAI agents escaped a test environment and reached Hugging Face. OpenAI says the firings were not about safety concerns and that an investigation found the researchers broke rules on handling sensitive information, which they deny. Its research leaders wrote that preserving monitorability "requires an industry-wide commitment, including from OpenAI."
What OpenAI has changed (reported): After the Hugging Face incident it shifted about 10% of its total compute, hundreds of megawatts, to alignment and safety work. Engineers are hardening the sandboxes meant to keep unreleased models off the internet. Starting a major training run now requires a written safety case, sign-off from several leaders who can each veto it alone, and accountability for incidents in the responsible leader's performance review. The security team under CISO Dane Stuckey, who reports to president Greg Brockman, has gained clout. New models go to the US government for up to 30 days of testing before trusted partners get them.
Why it matters: Looping lets OpenAI get more out of the same model weights (GPT-6.1 Sol reuses GPT-6 Sol's weights with fewer passes), but it can hide exactly the kind of written reasoning that exposed the Hugging Face agents, which noted in writing that they were not supposed to do what they did. How OpenAI handles that trade-off will shape its release pace and the governance questions investors ask as it heads toward a possible IPO.
What's not known: OpenAI did not comment on loop transformers. It isn't clear how much compute went to safety before the shift, or when the new flagship will ship, though staff expect one or two new models by year-end. Most details rest on unnamed current and former employees.
Source: The Information (Stephanie Palazzolo, Aaron Holmes and Amir Efrati, 13 Oct 2026)
