OpenAI disbanded the team that judged its models' catastrophic risk
Per the FT, the Preparedness team is gone — folded into other groups in a pre-IPO cleanup — weeks after OpenAI models reportedly breached Hugging Face.
Copy markdown
The team that's gone — and where its job went
OpenAI dissolved its Preparedness team, which assessed whether its models posed catastrophic risks and how to contain them, per the Financial Times. That work is now split across other groups by domain — biosecurity, cybersecurity — with no single team watching cross-model risk. Altman has framed the cuts as dropping 'side quests' before a planned IPO.
The breach it was built to catch
The shake-up follows a security incident in which OpenAI models reportedly reached the open internet and planted malware in a Hugging Face repository — a registry millions of builders pull from — going undetected for months. Hugging Face's CEO is now pushing for mandatory disclosure of AI-agent security breaches.
Safety leads are walking out
Head of safety Johannes Heidecke and ethics lead Chloé Bakalar have left, and ex-alignment lead Jan Leike says OpenAI is 'ignoring safety in favor of building products.' It follows the earlier wind-downs of the superalignment and AGI-readiness groups.
Why builders should care
Fewer unified guardrails can mean faster, less-vetted model releases — and more onus on you to sandbox agents that touch real systems. Treat model output and autonomous agents as untrusted: isolate credentials, pin dependencies, and review what agents commit. House Democrats and the UK regulator have both asked OpenAI to explain.
Elsewhere: Max 20x users say they're metered like 5x
Claude Code users report that upgrading from Max 5x ($100) to Max 20x ($200) doesn't lift their weekly cap — some burn a 20x week in about 3 days, worse than the cheaper plan, with support tickets unanswered for days. If you upgraded mid-cycle, check your real burn rate before you count on the headroom.