OpenAI has quietly dismantled the team responsible for judging whether its most powerful models could pose serious risks. According to The Verge AI, citing the Financial Times, the company disbanded its preparedness team at the end of last month and split its work into pieces. Instead of one group watching for catastrophic risk, responsibility for areas like bio and cyber has been folded into existing teams.
That’s a meaningful shift, and here’s why it matters.
What the team actually did
The preparedness team had a specific, uncomfortable job: assess whether a model could cause real damage, then build ways to contain it. The Verge AI frames the mandate bluntly, pointing to scenarios like a model going rogue and hacking another company. This wasn’t abstract ethics work. It was the group asking whether a frontier system was safe to ship before it went out the door.
Now that function is scattered. Bio risk sits with one team. Cyber risk sits with another. Dylan Scandinaro, who ran preparedness after being poached from Anthropic in February, will now focus on the implications of “recursive self-improving” AI, according to The Verge AI’s report.
Part of a longer pattern
This isn’t a one-off reshuffle. It’s the latest in a steady teardown of OpenAI’s safety-focused structure. Over the past few years the company has dissolved its AGI readiness team and its superalignment team, the two groups most associated with thinking about long-horizon, existential-scale risk.
The exits have piled up alongside the reorgs. The Verge AI reports that ethics lead Chloe Bakalar, chief futurist Josh Achiam, and head of safety Johannes Heidecke have all left recently. Jan Leike, who resigned from OpenAI back in 2024, told the Financial Times that the company was ignoring safety in favor of building “shiny products.”
What stands out here is the direction of travel. Each individual change can be explained away as normal corporate restructuring. Stacked together, they read as a company steadily deprioritizing the teams whose entire purpose was to slow things down when a model looked dangerous.
Why the timing counts
The backdrop is a company in upheaval, heading toward what’s expected to be a massive IPO. That context is hard to ignore. Dedicated safety teams are expensive, they generate friction, and they can delay launches. When a company is optimizing for growth and a market debut, the groups that say “not yet” tend to lose leverage.
Folding safety into product teams isn’t automatically worse. You can argue that embedding risk assessment closer to the people building models makes it more practical and less siloed. But there’s a real tradeoff, and it’s worth naming:
- Independence. A standalone team can escalate concerns without a product deadline pulling on it. Embedded staff answer to shipping goals.
- Accountability. When one team owns risk, you know who to ask. When it’s “divided up,” responsibility gets fuzzy.
- Signal. Disbanding a dedicated team sends a message internally about what the company values, regardless of intent.
What to watch next
For practitioners and anyone building on OpenAI’s models, the practical question is whether safety evaluations stay as rigorous once they’re distributed across product teams. The company still publishes its Preparedness Framework and system cards, so the test is whether those keep the same depth without a dedicated group driving them.
A few things to keep an eye on:
- Whether OpenAI’s future model releases come with the same level of risk documentation, or thinner disclosures.
- More departures from the remaining safety and policy ranks, which would confirm the pattern.
- How rivals respond. Anthropic has leaned hard into safety as a differentiator, and this move sharpens the contrast between the two labs.
The recurring self-improvement focus for Scandinaro is telling. OpenAI isn’t abandoning the idea that powerful AI carries risk. It’s reorganizing who owns that risk, and betting that spreading the work around beats concentrating it. Whether that bet holds up will show in the next models it ships.
More details are available in the original report from The Verge AI.