OpenAI Reorganizes Safety Work as IPO Pressure Builds

OpenAI has reportedly dissolved its preparedness team, the group responsible for stress-testing advanced models for severe risks, and redistributed that work across more specialized safety teams. The move, reported by the Financial Times and tied to the end of July, underscores how the company is reshaping its safety structure as it moves closer to a potentially huge public offering.

The reorganization matters because the preparedness team was designed to identify and reduce dangers from frontier AI systems, including cyber abuse, biosecurity threats, and other high-impact misuse scenarios. Its breakup adds to a broader pattern of safety-team upheaval at OpenAI, prompting fresh questions about whether the company is reducing centralized oversight at the same time it is racing to commercialize its products.

OpenAI has not publicly framed the change as a retreat from safety, but the timing has drawn scrutiny. Over the past several years, the company has repeatedly reworked or dissolved internal groups built to anticipate the long-term risks of artificial intelligence, while criticism from former staff and outside observers has intensified as the firm shifts toward consumer products, enterprise sales, and an anticipated IPO.

What happened to OpenAI’s preparedness team?

The preparedness team was disbanded at the end of last month, according to the Financial Times. Its responsibilities did not disappear, but the work was broken apart and assigned to teams focused on specific risk categories such as biological safety and cybersecurity.

That means OpenAI’s risk assessment function now appears to be more distributed than centralized. Instead of one group watching for a wide range of catastrophic or frontier-model dangers, the company has moved to a structure in which domain-specific teams handle their own slices of the safety problem.

In practice, that kind of reorganization can improve accountability in some areas, but it can also create gaps when threats overlap. A model capability that appears to be “just” a cyber issue, for example, may also have implications for fraud, privacy, and biosecurity.

How the team’s role fit into OpenAI’s safety system

The preparedness team was meant to serve as a kind of early-warning unit. Its job was to evaluate whether new models could pose severe harm and to help design ways to reduce those risks before a release.

That role became especially important as frontier models grew more capable, more autonomous, and more useful for a wider range of tasks. The possibility that a system could help an attacker automate phishing, write malicious code, assist with lab procedures, or even execute harmful instructions across software environments made preparedness testing a core part of the safety conversation.

OpenAI’s move suggests that responsibility for those judgments will now be embedded more deeply inside the company’s regular product and research teams rather than housed in a standalone unit.

Why does this matter as OpenAI heads toward an IPO?

The change matters because OpenAI is undergoing a major corporate transition while facing growing pressure to scale quickly. The company is widely expected to pursue a blockbuster initial public offering, and that prospect has intensified focus on how it balances revenue growth, product velocity, and safety governance.

Public markets tend to reward predictable execution, rapid growth, and clearer lines of accountability. Safety organizations, by contrast, are often designed to slow or block release decisions when models appear too risky. Those tensions become harder to manage when a company is moving from a quasi-research posture to a high-stakes commercial environment.

The latest restructure also arrives after years of leadership churn in the company’s safety-oriented units. Critics argue that OpenAI is increasingly prioritizing polished consumer and enterprise products over the more cautious, research-heavy posture it once emphasized.

Former OpenAI researcher Jan Leike, who left the company in 2024, has argued that the company appears to be favoring “shiny products” over the kind of deep safety work he считает essential for frontier AI development, according to the Financial Times.

That criticism reflects a broader debate throughout the AI industry: whether the race to ship useful tools is outpacing the institutions built to manage their risks.

How OpenAI’s safety structure has changed over time

OpenAI has repeatedly reorganized its safety teams in ways that reduce standalone oversight. The preparedness team is only the latest in a series of changes that have altered how the company handles long-term risk.

In recent years, the company has dissolved its AGI readiness group and its superalignment team, both of which were linked to attempts to anticipate and constrain the most extreme capabilities of future systems. Those moves have fueled concern that OpenAI is gradually replacing broad, centralized safety planning with narrower, more product-aligned processes.

At the same time, several prominent safety figures have departed or shifted roles. The recent exits of ethics lead Chloé Bakalar, Chief Futurist Josh Achiam, and safety head Johannes Heidecke have deepened the impression of a team in transition.

Who is taking over the work now?

The work has reportedly been split across existing teams, especially in biosecurity and cyber safety. That structure means specialized groups now carry responsibility for the areas previously coordinated by the preparedness team.

OpenAI’s preparedness lead, Dylan Scandinaro, who joined from Anthropic in February, is said to be refocusing on the implications of “recursive self-improving” AI. That term generally refers to the idea that a system could improve its own capabilities, potentially accelerating progress faster than human institutions can monitor.

It is a particularly sensitive area of research because recursive self-improvement raises difficult questions about control, oversight, and whether safety systems can keep up with rapidly advancing models.

What is recursive self-improving AI, and why is OpenAI still watching it?

Recursive self-improving AI is a system that can help improve itself, and that makes it one of the most closely watched frontier risks. If an AI can meaningfully accelerate its own development, it could become harder to predict, govern, or safely contain.

While the concept remains theoretical in many respects, it sits near the center of long-running debates about whether advanced AI could eventually become difficult for humans to control. That is why OpenAI appears to be keeping a dedicated eye on the issue even as its broader preparedness function is dismantled.

The fact that a senior safety leader is being reassigned to that topic suggests OpenAI has not abandoned long-horizon risk work entirely. But it also indicates that the company is narrowing where that work lives and how visible it is inside the organization.

Timeline of OpenAI’s safety shakeups

OpenAI’s latest move fits a wider sequence of internal reorganizations. The company has altered its safety architecture several times as its models have become more powerful and its commercial ambitions have expanded.

This sequence does not prove that OpenAI is abandoning safety. It does, however, show a consistent preference for folding risk work into operational teams rather than maintaining large standalone units focused on far-future scenarios.

Why are critics worried?

Critics worry that OpenAI’s safety culture may be losing independence as the company grows more commercial. When safety teams become smaller, less central, or more closely tied to product groups, it can be harder for them to challenge release timelines or flag uncomfortable findings.

That concern is especially acute for frontier AI companies because the stakes are unusually high. The same model that powers a useful assistant can also help with malware generation, sophisticated social engineering, synthetic biology workflows, or large-scale deception if misused.

For outside observers, the question is not whether OpenAI has safety staff. It does. The question is whether those staff have enough authority, visibility, and insulation from growth pressure to slow down systems that may be too risky to deploy.

What supporters might say in response

Supporters could argue that specialization makes the safety function more effective. Domain experts working inside dedicated cyber or bio teams may be better equipped to evaluate specific threats than a smaller centralized unit trying to cover everything at once.

They might also point out that incorporating risk reviews into product and research teams can make safety a routine part of development rather than a separate compliance exercise. In that view, the reorganization could reflect maturity, not retreat.

Still, the burden remains on OpenAI to show that the new setup is at least as robust as the one it replaced.

How does this compare with other AI companies?

OpenAI is not alone in wrestling with how to structure AI safety, but it is one of the most scrutinized companies in the field. Other major developers have built dedicated policy, trust, and safety organizations, though the size and authority of those teams vary widely.

The difference at OpenAI is the combination of scale, visibility, and timing. The company sits near the center of the generative AI boom, its products are used by consumers and businesses around the world, and its internal changes are watched closely because they can influence broader industry norms.

When a company this influential reduces the prominence of standalone safety units, the move can ripple beyond its walls. Competitors, regulators, investors, and researchers all take signals from how OpenAI chooses to organize itself.

What should investors and regulators watch next?

Investors and regulators should watch whether OpenAI can document stronger governance after the restructure. The key issue is not just who reports to whom, but whether the company can prove that dangerous capabilities are still being tested, challenged, and, when necessary, delayed.

Several questions will likely matter most in the months ahead:

• Will OpenAI publish clearer safety benchmarks or evaluation practices?
• Will specialized teams have authority to stop launches?
• How will the company handle overlapping risks across cyber, bio, and model autonomy?
• Will departures from safety leadership continue?
• Can OpenAI reassure policymakers as IPO preparations intensify?

Those answers will shape whether this restructuring is remembered as an efficiency move or as another milestone in the steady erosion of independent safety oversight.

The bigger picture

OpenAI’s preparedness-team shutdown is part of a larger debate over whether AI companies can police themselves while racing to dominate the market. The company has become a symbol of that tension: simultaneously a frontier research lab, a consumer-tech giant, an enterprise software provider, and a potential public company with enormous financial upside.

As models become more powerful, safety work becomes more consequential, not less. That is why changes to internal oversight structures attract so much attention. They are not just organizational details; they are clues about what a company believes matters most as it scales.

If OpenAI can show that specialized safety teams are more effective than a centralized preparedness unit, the reorganization may fade into the background. If not, it could become another example of the company trimming the very systems designed to keep its most advanced AI under control.

For now, the message is clear: OpenAI is still moving fast, and its approach to safety is changing just as rapidly.