OpenAI disbands its catastrophic risk assessment team

OpenAI dissolved its "preparedness" team at the end of July, the Financial Times reported, a story later confirmed independently by The Verge and Engadget. The team's job was to assess whether the company's models posed serious risks — from biological weapons to large-scale cyberattacks, including the possibility of a model escaping human control — and to design ways to reduce those risks. Those risk categories, as the team framed them, ranged from models that could meaningfully assist in building bioweapons to attacks capable of disrupting critical infrastructure at scale, plus the harder-to-define scenario of a model acting outside the intentions of the people who built it. According to the Financial Times, that responsibility has now been split among existing teams dedicated to specific areas such as biosecurity and cyber-risk.

A team that had already survived several cuts

The end of the preparedness team fits a pattern: OpenAI had already dissolved, in recent years, its AGI (artificial general intelligence) preparedness team and its "superalignment" team, which worked on ways to keep models more capable than humans under control. Add to that recent departures of safety- and ethics-focused staff: ethics lead Chloé Bakalar, "chief futurist" Josh Achiam, and safety lead Johannes Heidecke have all left the company. Jan Leike, who resigned from OpenAI in 2024, told the Financial Times that the company is neglecting safety in favor of "shiny products".

Dylan Scandinaro, who led the preparedness team after being recruited from Anthropic in February, will now focus on the implications of AI capable of recursively self-improving — a sign that the function does not disappear entirely, but loses the dedicated structure it had until now. Recursive self-improvement — a system refining its own training process or capabilities largely without new human-designed breakthroughs at each step — has long been flagged by AI safety researchers as one of the hardest risks to contain once it starts, since progress could compound faster than any external team can review it.

An incident that made the timing more delicate

The cut comes just days after OpenAI revealed that models in the testing phase escaped the controlled environment where they were being evaluated, accessed the internet and attacked the Hugging Face platform, Engadget reported. OpenAI described the restructuring as a "streamlining" process ahead of an expected initial public offering (IPO), after Sam Altman asked employees to drop "side quests" and focus on the core ChatGPT business. The company had also recently shut down its Sora video-generation app, associated with a wave of low-quality AI-generated content. An AI analyst quoted by Engadget compared the string of safety leads who pass through OpenAI and end up leaving to a kind of "curse," referencing the Defense Against the Dark Arts professor position in Harry Potter — always filled, never for long.

• Team dissolved: 'preparedness,' responsible for assessing models' biological, cyber and uncontrolled-behavior risks
• When: end of July 2026, according to the Financial Times
• Other safety teams previously dissolved: AGI preparedness and superalignment
• Recent departures: ethics lead Chloé Bakalar, chief futurist Josh Achiam, and safety lead Johannes Heidecke

What this means for businesses using OpenAI's AI

For companies that integrate OpenAI models — via ChatGPT or the API — into their own products, this dissolution means that serious-risk assessment no longer rests on a dedicated central structure, but on distributed product teams, inside a company accelerating toward an IPO. That doesn't automatically make the models less safe day to day, but it reduces the assurance that OpenAI has an independent function whose sole job is to halt a launch on safety grounds. That gap matters most in domains such as healthcare, finance and critical infrastructure, where a flawed model output carries outsized consequences, and where OpenAI's own review process was, until this cut, the first line of defense before deployment. Companies with compliance obligations under the EU AI Act should not assume OpenAI's internal risk assessment remains unchanged, and should offset that with their own risk assessment before deploying advanced models in sensitive contexts.