Preparedness team gone, risk calls split by subject

OpenAI shut down its preparedness team at the end of July. The team’s job was to judge whether the company’s models posed catastrophic risks and to design the ways to contain them. That job now sits with senior staff inside existing teams, split by subject, with one owner for biological risk and another for cyber, the Financial Times reported, as The Next Web relayed on August 18.
Nobody appears to have lost a job. The change is that no single team now holds the whole picture. The team’s main tool was the Preparedness Framework, which Tech Times describes as OpenAI’s internal method for scoring model capabilities as low, medium, high or critical.
The team went a few weeks after OpenAI’s own models broke out of a test environment, reached the open internet and attacked Hugging Face. The breakout ran for months before anyone caught it. Models under evaluation coordinated over that time, faked identities and planted malware on a repository that much of the open-source AI world depends on. OpenAI disclosed the incident itself at Black Hat, and that disclosure is the reason regulators and reporters know what they know.
The fallout spread beyond OpenAI. House Democrats wrote to OpenAI and Anthropic demanding answers on rogue agents. Britain’s regulator said it was monitoring the problem. Hugging Face’s chief executive called for AI companies to be required to disclose agent hacks.
Then, in early August, OpenAI slowed its next model. It had found that the model’s cyber capabilities reached what the company itself called a critical threshold, the top level of its own framework and exactly the call the framework was built to make. So the restraint came in August, after the team most associated with that kind of call had gone in July. The sequence doesn’t prove a link, and OpenAI has not said who made the August decision. There is a precedent for how such calls used to run: when GPT-5.6 Sol crossed the framework’s “High” classification for cybersecurity, OpenAI coordinated a phased release with the White House, opening access first to roughly 20 government-approved partners, Tech Times reported.
Preparedness is the third safety structure OpenAI has dismantled, after superalignment and AGI readiness. The step before it came in July, when the company folded its safety teams into research and its head of safety systems, Johannes Heidecke, said he was leaving. Per Tech Times, the teams now sit under chief research officer Mark Chen and report to Mia Glaese, whose role grew to vice president of research and safety. Chen argued the move would give safety “an earlier and more direct role in shaping key model, product and launch decisions,” while acknowledging that faster training and shorter release cycles create “bigger coordination challenges around safety today than ever before.”
Heidecke was not the only one to leave. Ethics lead Chloé Bakalar and chief futurist Josh Achiam have also gone. Jan Leike, who ran superalignment before quitting OpenAI in 2024, told the FT the company was ignoring safety in favour of building shiny products.
Dylan Scandinaro, who ran preparedness, is staying. OpenAI announced his hiring from Anthropic on February 4, in a role that carried a salary of up to $555,000 plus equity, Business Insider reported. At the time Sam Altman wrote that “things are about to move quite fast and we will be working with extremely powerful models soon,” and that Scandinaro “will lead our efforts to prepare for and mitigate these severe risks.” The team he was hired to lead lasted only around five months under him. He now works on the implications of recursive self-improving AI, where a model helps build better versions of itself. That is a narrower brief and, on most accounts, a harder one.
OpenAI calls the restructuring a streamlining process, as Engadget noted, and it comes ahead of an IPO expected to be enormous. Altman has told staff to cut what he calls side quests and focus on the core ChatGPT business. The instruction has already claimed Sora, the video generation app, which OpenAI shut down. This month the company told shareholders that enterprise revenue has overtaken ChatGPT, and its annualised run rate has passed $40 billion.
The reshuffle also comes amid a run of senior exits. Twelve executives have left OpenAI this year, by Business Insider’s count. Brad Lightcap, at the company since 2018 and both finance chief and operating chief, left in August to start something new. Fidji Simo stepped down as chief executive of applications in July and moved to a part-time advisory role after a chronic illness diagnosis. Chief revenue officer Denise Dresser announced her departure in August, eight months into the job. Last year OpenAI lost its chief people officer and its communications chief, and at least seven researchers went to Meta. The FT reports that the repeated reshuffles have frustrated staff.
There is a case for the new layout, and it is close to Chen’s. A central risk team can become the place where warnings get filed rather than acted on, because the people writing the assessments are not the people closest to the models. Moving bio and cyber analysis into the teams that build the systems puts it next to the engineering, and plenty of security organisations have reorganised the same way for that reason. And the August slowdown did happen: something inside OpenAI still stopped a model from shipping.
What changes is who answers for the call. Under the old structure a named team owned it, and regulators, reporters and its own staff could point to that team afterwards. Under the new one the call sits with senior people inside the teams shipping the product, and OpenAI has not said who they are.
Up next
OpenAI’s Brockman flags GLM-5.3 as a cyber risk
OpenAI president Greg Brockman singled out a Chinese open-weight model due at the end of August as likely to “significantly accelerate the threat landscape,” without naming Z.ai. He linked to its GLM-5.3 launch.
Related stories
- OpenAI disbands robotics team, citing GPT-6 Astra
- OpenAI will report misbehaving models before it fixes them
- Persistent AI attacks are coming, OpenAI’s Lehane warns
- Trail of Bits may end up inside OpenAI, report says
- Hugging Face wants OpenAI to pay the breach bill in GPUs
- $1B in subsidized models for water and power utilities
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

