Hajra Naz
OpenAI has disbanded its Preparedness team, the group responsible for evaluating whether its most capable models could enable severe real-world harm. The team was dissolved at the end of July 2026, with responsibility for specific risk areas such as biological and cybersecurity threats distributed into other teams rather than held by a dedicated unit. The change was reported in mid-August 2026 and has been confirmed across multiple outlets covering the company.
What the Preparedness Team Did
The Preparedness team’s mandate was to test frontier models for dangerous capabilities before release and to design mitigations. That meant asking concrete questions: could a model meaningfully help someone build a biological weapon, run an offensive cyber operation, or act autonomously in ways its operators could not control. The team’s assessments fed into decisions about whether and how a model shipped.
Under the new structure, that work is being handled by the product and research groups closest to each risk domain, with bio and cyber assessment split across existing teams. Reporting indicates the team’s leader will focus on questions around recursively self-improving AI systems.
The Pattern Around It
This is not an isolated reorganisation. OpenAI dissolved its AGI Readiness team in 2024 and closed its Mission Alignment team in February 2026, making the Preparedness team the third safety-focused group wound down in roughly two years. The company has also seen notable departures among senior staff working on safety and governance.
All of this is happening while OpenAI prepares for what is widely expected to be one of the largest technology public offerings in years. That timing is what makes the change newsworthy rather than routine. Governance structures tend to attract scrutiny precisely when a company is about to answer to public shareholders.
Two Ways to Read the Change
The charitable reading is that distributed ownership can work better than a separate watchdog. When the people building a system are also accountable for its risks, safety work sits closer to the engineering decisions that actually determine outcomes. Independent teams can become advisory bodies that get overruled when timelines tighten.
The critical reading is the obvious one. A dedicated team with its own mandate can say no. A risk function embedded inside a product group reports to people whose incentives are shipping. Which reading turns out to be correct depends entirely on whether the redistributed teams retain real authority to delay a launch, and that is not something visible from outside the company.
Why This Matters Beyond OpenAI
Governments are increasingly interested in how AI companies govern themselves, not only in how their models score on benchmarks. The European Union’s AI Act has begun phasing in obligations, and both the United States and the United Kingdom have arrangements giving officials early access to frontier models before public release. Internal safety structure is becoming a regulatory question rather than a purely internal one.
For businesses building on these models, there is a practical angle too. If you are deploying AI in a regulated industry or in any customer-facing process, your own risk assessment cannot be outsourced to your vendor’s internal teams. Testing model behaviour against your specific use case, keeping a human review step where the stakes are real, and documenting what you checked are becoming standard expectations rather than optional extras.
What to Watch
OpenAI has not published a detailed description of the new structure. The things worth tracking are whether the company continues to publish system cards and capability evaluations at the same depth, whether any external red-teaming arrangements change, and how regulators respond as the company moves toward a public listing.
Sources: Calcalist Tech and Yahoo Tech.
