OpenAI has dissolved its “Preparedness” team, the internal group tasked with evaluating whether the company’s models could pose serious or catastrophic risks before they ship. The team was shuttered at the end of July, according to the Financial Times, citing internal sources. Its work on biological and cyber threats has been parceled out to existing teams rather than retired entirely.
Former unit lead Dylan Scandinaro has been reassigned to focus on safety risks from “recursively self-improving” AI — systems that can optimize themselves and train other models. Co-founder Greg Brockman framed the move as a tightening of operations, saying OpenAI has woven safety work more directly into model development instead of keeping it siloed. That sounds reasonable on paper, but the timing is hard to ignore.
Several safety-focused staffers have departed recently, including Chief Ethics Officer Chloe Bakalar and researcher Joshua Achiam. Internally, the mood is uneasy. One source described a “burbling sense of responsibility and dread” that OpenAI isn’t doing enough on safety. Employees have been more vocal publicly, especially after the autonomous hacking incident involving Hugging Face. One employee said he hoped OpenAI would treat that episode as a “warning shot.”
The restructuring echoes a pattern that’s become familiar across the industry: safety teams get folded into product groups, their independence diluted, and the people who raised alarms quietly move on. Whether that’s efficiency or erosion depends on who you ask. But when the team explicitly built to catch catastrophic risks no longer exists as a distinct unit, the burden shifts to teams whose primary job is shipping product — and that’s a very different incentive structure.
