OpenAI has quietly shut down the team it built to watch for the most dangerous things its own AI might do. According to a Financial Times report, the company dissolved its centralized Preparedness team, the group responsible for judging whether its models could help with catastrophic harms such as the development of biological weapons or large scale cyberattacks, and the change matters because that team existed precisely to flag the rare, worst case failures that ordinary product testing is not designed to catch.
The work has not vanished so much as been scattered, since OpenAI has reassigned senior members of the group to existing teams that now own preparedness in their own domains, with responsibility for cyber and bio risks handed to the units already working in those areas. The company frames the reorganization as streamlining, an argument that folding safety expertise directly into the teams building and shipping the technology puts that expertise closer to the decisions that matter rather than leaving it in a separate silo.
Critics read the same facts differently, and the reason is history, because this is the third safety focused unit that OpenAI has dissolved in roughly two years, following earlier teams that were also broken up or absorbed after prominent researchers left. When a company repeatedly stands up a dedicated safety group and then disbands it, the pattern invites the worry that central oversight keeps losing ground to the pressure to move quickly, whatever the stated rationale for any single change.
The timing sharpens that worry, since the reorganization arrives during a stretch of high profile departures and internal reshuffling as OpenAI prepares for an expected public offering, a moment when the incentives to present a lean, fast moving organization are especially strong. A team whose entire job is to slow things down when a model looks too capable in a dangerous direction is, by design, in tension with that momentum, which is exactly why observers watch what happens to such groups so closely.
None of this means OpenAI has abandoned the work, and the company is clear that preparedness continues inside the domain teams rather than ending, so the honest reading is not that the guardrails are gone but that they have moved. The open question is whether catastrophic risk evaluation carries the same weight when it is one responsibility among many inside a product team as it did when a dedicated group answered for it alone. That is a difficult thing to measure from the outside, and it stands in contrast to the path taken by rivals such as Anthropic, which recently published its own risk assessment and raised the alarm on its models in public. This is a sensitive area, and reasonable people inside and outside the company will disagree about whether distributing the work strengthens it or dilutes it, a judgment the next serious safety test will help settle.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →