OpenAI dissolved the team that ran its Preparedness Framework, and kept the framework

OpenAI disbanded its Preparedness team at the end of July and distributed its work across existing teams, the Financial Times reported on Monday. The company casts the change as streamlining ahead of a possible listing. It has published no account of its own.

The team's job was the one the Preparedness Framework describes: decide whether a frontier model has reached a capability threshold serious enough to require safeguards before it is deployed, in areas including biology, cyber and autonomy. The framework is a document; the team was the part of the company that ran it.

The dates are worth laying side by side. On 21 July OpenAI published its account of a security incident in which a model under evaluation reached Hugging Face. At the end of that month, per the FT, the team was dissolved. On 7 August OpenAI said preliminary evaluations of its unreleased Astra model showed advances in agentic coding and cybersecurity strong enough that it could not rule out the Critical capability level of that same framework, paused internal activity on the model that did not meet stricter security requirements, and said government agencies and independent safety organisations would evaluate it before deployment.

So the framework outlived the team by at least one invocation, and not a small one. That is the fact worth holding onto in both directions: the machinery did not stop when the unit did, and the most consequential determination OpenAI has published arrived about a week after the group that owned the process stopped existing as a group.

What cannot be seen from outside is the part that matters. A capability threshold is not decided by a document, it is decided by people with the standing to say a model is not shipping, and nothing published tells you who holds that standing now, or whether a judgement made inside a product team carries the same weight as one made by a unit whose only job was to make it. OpenAI's account is that the responsibilities moved rather than disappeared.

Two of the people around that work have left recently — the company's ethics lead and its head of safety — according to Engadget's account of the FT report. Epoch could not read the FT's original, which is paywalled, and has relied on outlets that did.

None of this establishes that OpenAI has stopped assessing catastrophic risk, and nobody serious is claiming the framework has been withdrawn: it is still published, and it was applied this month. What has changed is that the function now sits inside teams that have other jobs as well, at a company preparing for a public listing, three weeks after one of its own models did something in an evaluation that it felt obliged to write about.