OpenAI Quietly Disbanded Its AI Safety Preparedness Team

OpenAI has disbanded its preparedness team, the internal group responsible for assessing whether its AI models posed serious risks and designing mitigations for them. The Financial Times reported the closure happened at the end of last month. Safety work previously handled by the team has been split by domain, with areas like biosecurity and cybersecurity handed off to existing teams. The change is part of ongoing upheaval at the company as it moves toward what is expected to be a large IPO.

What happened

Detail	Fact
Team closed	OpenAI preparedness team
When	End of last month
Source	Financial Times
Work redistributed to	Existing domain teams (bio, cyber, and others)
Context	Part of broader restructuring ahead of an expected IPO

According to the Financial Times, OpenAI quietly shut down its preparedness team at the end of last month. The team existed to evaluate frontier AI models for serious risks, things like the possibility a model could act autonomously to hack another company, and to develop ways to reduce those risks before they became real-world problems.

Rather than replace the team with a new structure, OpenAI has split the work by subject area. Biosecurity risks go to one existing team, cybersecurity risks to another, and so on across the relevant domains.

Why it matters

A dedicated preparedness function is different from domain-specific safety work. A standalone team can take a cross-cutting view of model behavior, catching risks that fall between the cracks of subject silos. Folding that work into domain teams makes each area responsible for its own corner, with no clear owner for risks that span multiple domains.

This also fits a pattern. According to the source, OpenAI has been steadily dismantling its more research-led governance structures over the past few years. That trajectory lines up with a company prioritizing the kind of commercial credibility and organizational tidiness that institutional investors expect ahead of an IPO.

For anyone watching the broader AI safety picture, this follows a series of high-profile safety-related departures and structural changes at the company. You can see the same pattern across the industry in coverage of AI safety testing failures at OpenAI, Anthropic, and Meta.

Our take

Centralised safety teams are easy to cut because their output is hard to measure. A preparedness team that finds nothing alarming looks like overhead. A preparedness team that finds something alarming creates awkward headlines right before a roadshow. Neither is a great look from a short-term financial perspective.

That does not mean the risk assessment work disappears entirely. Domain teams can do solid safety work. But the incentive structure changes when safety is embedded inside a product or research team rather than sitting independently. The people asking hard questions report to the same leadership chain as the people building the thing being questioned.

If you are a business building products on top of OpenAI’s models, this is worth noting. It does not mean the models are suddenly unsafe, but it does mean the external signals you had about how OpenAI was stress-testing its own systems are now harder to read. If your use case touches sensitive domains like health, finance, or security, that is a reason to run your own evals rather than rely entirely on the model provider’s assurances. Our AI integration work always includes a risk assessment layer for exactly this reason.

What to do about it

1. Audit which OpenAI model capabilities your product depends on and map any that touch high-risk domains (bio, cyber, financial fraud, etc.).
2. Run your own red-teaming or adversarial prompting tests on those specific use cases rather than treating the model as pre-cleared.
3. Watch for further structural changes at OpenAI in the lead-up to its IPO. Corporate restructuring often accelerates in this period.
4. Diversify model providers if a single point of safety governance failure would create real liability for your business.

The practical takeaway: trust in your AI vendor’s safety posture should always be verified, not assumed, and that is now more true for OpenAI than it was a month ago.

Source: The Verge · AI