OpenAI Reportedly Disbanded Its Preparedness Team

OpenAI quietly disbanded its Preparedness team at the end of July 2026, according to reporting from the Financial Times and The Verge — making it the third dedicated AI safety group the company has dissolved in roughly two years. The team's mandate had been to answer some of the most consequential questions in AI development: could this model be used to engineer a bioweapon, conduct unauthorized cyberattacks on other organizations, or act autonomously in ways its creators never intended?

Background You Need

The Preparedness team was created to sit outside the normal product development loop and ask pre-deployment questions that product teams might be structurally incentivized to downplay. Its job wasn't to make models smarter or faster — it was to determine whether a given model crossed thresholds that made it too dangerous to ship in its current form. That kind of adversarial internal review is expensive, slow, and resistant to the competitive pressure to release. It is also, safety researchers would argue, exactly what separates responsible deployment from a race to the bottom.

OpenAI has been here before — twice. The AGI Readiness team was disbanded in 2024. Then, in February 2026, the Mission Alignment team was dissolved, its leader elevated to a newly created "chief futurist" role that multiple observers noted carried no operational responsibility for safety outcomes. The Tech Buzz, which covered that February disbandment, described the pattern succinctly: responsibilities get folded into everyone's job description, which often means they become no one's priority. Each time, the official explanation has been some version of the same claim — the work isn't disappearing, it's being integrated more deeply into the organization. The Preparedness team's dissolution in July comes with an identical framing.

The timing is not incidental. Crypto Briefing and The Verge both flag OpenAI's expected IPO as an explicit driver. A company preparing to go public has structural incentives to smooth out internal friction, rationalize headcount, and demonstrate operational efficiency — goals that are in direct tension with maintaining a team whose entire purpose is to pump the brakes.

What's New

According to The Verge's reporting, the Preparedness team's responsibilities have been split by domain: biological risk assessment and cybersecurity risk assessment now live in separate existing teams. On paper, the coverage is the same. In practice, those assessments are now being conducted by teams that also have product delivery obligations — a structural conflict that is difficult to dismiss as cosmetic.

Dylan Scandinaro, who led the Preparedness team after being recruited from Anthropic in February 2026, has not left the company. His mandate has narrowed significantly, however. Multiple reports indicate he is now focused specifically on the problem of recursive self-improvement — the scenario in which an AI system enhances its own capabilities without human oversight. That is a meaningful and legitimate area of research, but it is a subset of what the Preparedness team covered. The broader threat surface — cyberweapons, biological risk, unauthorized autonomous action — has been redistributed outward.

The personnel losses compounding this restructuring are harder to explain away as simple organizational streamlining. Chloé Bakalar, OpenAI's ethics lead, departed recently. Josh Achiam, who held the chief futurist title created in February, has also left. Johannes Heidecke, who served as head of safety, is gone as well. Three people whose explicit job was to think about what could go wrong are no longer at the company that is building the systems in question.

Jan Leike, who resigned from OpenAI's superalignment team in 2024, went on record with the Financial Times characterizing the pattern as a company choosing "shiny products" over safety. His read at the time read as a frustrated insider's complaint. By August 2026, the data has accumulated to the point where it looks more like a documented trend than a disgruntled exit interview.

Our read is that the structure here matters as much as the stated intent. When a company says "safety is now everyone's responsibility," that is sometimes true and sometimes a way to dilute accountability without saying so directly. The test is what happens when a product team's safety assessment recommends delaying a release. In a dedicated safety group with independent standing, that recommendation carries organizational weight. Inside a product team, it competes with shipping timelines and quarterly targets. OpenAI has not explained how it plans to resolve that tension.

The Pushback

OpenAI co-founder Greg Brockman has argued, per Crypto Briefing's coverage, that embedding safety work directly into model development and product teams is actually a stronger approach than maintaining a siloed group — the theory being that safety becomes intrinsic to the process rather than an external checkpoint that can be bypassed or deprioritized. It is a defensible position in the abstract. Integrated safety culture is genuinely preferable to a compliance-style team that exists to tick a box. The problem is that the integrated approach requires trusting that product teams will surface and escalate findings that slow down their own work, without the structural backstop of an independent group empowered to do exactly that.

What the three consecutive dissolutions do not answer is what organizational mechanism now exists to tell OpenAI's leadership that a given model should not ship. A preparedness team, a mission alignment team, an AGI readiness team — each was that mechanism. All three are gone. The corroborating sources agree on the facts and diverge only in how charitably they interpret the intent: AIToolly frames it as a risk management transition; The Tech Buzz calls it a signal that safety has become "no one's priority." The structural reality supports the latter reading more than the former.

Sources