OpenAI has reportedly disbanded its preparedness team, the latest in a series of safety-focused reorganizations as the company barrels toward a highly anticipated initial public offering. According to the Financial Times, the team was dissolved at the end of last month, scattering its responsibilities across existing departments.
The preparedness team was originally tasked with evaluating whether OpenAI's models posed severe risks—such as the potential for an AI system to go rogue, hack into other companies, or enable bioterrorism—and developing mitigation strategies to address those dangers. Its dissolution marks a significant shift in how the company approaches frontier AI safety, particularly at a time when critics argue that the pressure to ship commercial products is eroding long-standing safety commitments.
From preparedness to distributed oversight
Following the disbandment, the responsibilities that once belonged to the preparedness team have reportedly been divided into specialized areas and moved into existing organizational units. For example, biological and chemical risk assessments will now be handled by teams that already focus on those domains, while cybersecurity risks will be absorbed into the company’s broader security and safety infrastructure. The decision suggests a move away from a centralized, dedicated preparedness function toward a more integrated but arguably less visible model of risk management.
The shift also raises questions about accountability. A dedicated preparedness team offered a clear focal point for safety concerns, with a mandate to report directly to senior leadership. Now, with those duties distributed, it may become harder for outside observers—and even internal employees—to track how thoroughly these risks are being evaluated before models are deployed.
A history of safety-team upheaval
This is not the first time OpenAI has reorganized its safety apparatus. Over the past few years, the company has gradually dismantled several high-profile safety initiatives. The AGI readiness team, which was designed to ensure alignment with the company’s mission as it approached artificial general intelligence, was dissolved. The superalignment team, once led by chief scientist Ilya Sutskever and dedicated to solving the problem of controlling superintelligent AI, was also disbanded.
Key figures have departed as well. Ethics lead Chloé Bakalar, Chief Futurist Josh Achiam, and head of safety Johannes Heidecke have all left the company in recent months. Their exits, along with the earlier resignation of alignment researcher Jan Leike in 2024, have fueled criticism that OpenAI is prioritizing commercial success over its founding principles.
Jan Leike’s warning
Jan Leike, who resigned from OpenAI in 2024 after leading the superalignment team, has been particularly outspoken. In interviews with the Financial Times, Leike accused the company of ignoring safety in favor of developing “shiny products.” He has argued that the pressure to release increasingly capable AI models—such as GPT-4o and successors—has come at the expense of rigorous safety testing and long-term risk mitigation.
Leike’s departure was not an isolated event. His resignation followed a period of internal turbulence at OpenAI, which included the dramatic ouster and reinstatement of CEO Sam Altman in late 2023. That episode highlighted deep divisions within the company about the speed of AI development and the adequacy of governance structures.
The role of Dylan Scandinaro
Dylan Scandinaro, the head of the now-disbanded preparedness team, was poached from rival AI lab Anthropic in February. Under the new arrangement, Scandinaro will reportedly focus on the implications of “recursive self-improving” AI—a concept where an AI system can improve its own code and capabilities in an iterative cycle, potentially leading to explosive intelligence growth. This is considered one of the most profound and least understood risks in frontier AI development.
Scandinaro’s background at Anthropic, a company known for its safety-focused approach, had been seen as a signal that OpenAI was serious about maintaining robust preparedness measures. His shift to a more specialized research role could be interpreted either as a strategic deepening of focus on the most dangerous long-term scenario, or as a move to sideline an independent safety voice.
Context: OpenAI’s march toward an IPO
The reorganisation comes as OpenAI is reportedly preparing for what is expected to be one of the largest tech IPOs in recent history. The company has been actively restructuring its corporate governance and financial operations to appeal to public market investors. A successful IPO would solidify OpenAI’s position as the dominant commercial AI lab, but it also brings new pressure to demonstrate profitability and growth.
Public market investors tend to reward speed and revenue, not precaution. If OpenAI is forced to justify every safety measure in terms of its impact on product releases, the incentive structure could further weaken the company’s already-frayed safety culture. Analysts have noted that the disbandment of the preparedness team, combined with the departure of senior safety leaders, could make the company more nimble in the short term but more vulnerable to catastrophic failures in the long term.
Industry reactions
The news has drawn sharp reactions from the AI safety research community. Some argue that OpenAI’s approach reflects a broader industry trend: as AI models become more commercially valuable, dedicated safety teams are often seen as obstacles rather than assets. Others point out that distributing risk assessment across domain-specific teams is not inherently wrong—it could even lead to more thorough, context-aware evaluations if properly implemented.
However, the lack of transparency about how the new structure will work has raised concerns. OpenAI has not publicly detailed the reporting lines for the newly distributed preparedness functions, nor has it clarified how independent oversight will be maintained. In previous statements, the company has insisted that safety remains its top priority, but the repeated dismantling of dedicated safety teams tells a different story to many observers.
Comparing with Anthropic and other labs
Anthropic, Scandinaro’s former employer, has historically maintained a more decentralized safety approach, with safety considerations integrated into every stage of model development. That model is often cited as an alternative to OpenAI’s standalone teams. But even Anthropic has faced criticisms from its own employees about insufficient safety practices.
Other major AI labs, such as Google DeepMind, have also restructured their safety divisions over the years. Yet few have done so as publicly and as rapidly as OpenAI. The constant changes at the company can make it difficult for regulators and researchers to keep track of who is responsible for what, and whether any given risk is being properly managed.
What does this mean for AI risk?
The dissolution of the preparedness team does not necessarily mean that AI risk assessments will be neglected. Many of the same experts will likely continue their work under different banners. But organisational changes of this kind can have subtle but significant effects on risk culture. When a dedicated team is disbanded, its members may be scattered across departments, losing their collective identity and institutional memory. Their recommendations may also be diluted by competing priorities within larger teams.
Moreover, the shift in focus toward “recursive self-improving” AI suggests that OpenAI is looking at the most extreme long-term speculative risk while potentially paying less attention to near-term harms. Those near-term harms include algorithmic bias, privacy violations, deepfakes, and the use of AI in disinformation campaigns—all of which have already been observed in the wild.
The company has faced several high-profile controversies in recent months, including reports about its ChatGPT product tracking user clicks and keystrokes, and debates about watermarking AI-generated content. While these issues are not directly related to the preparedness team’s mandate, they illustrate the broader tension between shipping features and ensuring safety.
A changing safety culture
OpenAI was founded in 2015 as a nonprofit with a mission to ensure that artificial general intelligence benefits all of humanity. In 2019, it created an unusual capped-profit structure to raise capital while ostensibly preserving its mission. Over time, that structure has come under increasing strain as the company has pursued larger commercial deals, including partnerships with Microsoft and a massive global consumer rollout.
Safety culture researcher Lotte Bailey, who has studied AI labs, notes that the departure of ethics and safety leaders often has a corrosive effect on the teams that remain. “When senior people leave, it sends a signal that safety is not a career path within the company,” Bailey said in a recent interview. “Those who stay may become less willing to raise concerns, fearing that they will be sidelined or let go.”
Dylan Scandinaro’s new focus on recursive self-improvement could be seen as an attempt to retain at least one high-profile researcher in a safety-adjacent role. But whether he will have the authority to halt or delay a model release if he identifies a serious risk remains unknown.
Regulatory implications
Governments around the world are still grappling with how to regulate AI. The European Union has implemented the AI Act, which imposes strict requirements on high-risk AI systems. The Biden administration in the United States issued an executive order on AI safety in late 2023, although subsequent administrations have taken a lighter-touch approach.
Regulators often rely on voluntary commitments from AI labs and on internal safety documents. If OpenAI’s internal safety structures become more fragmented, it may become increasingly difficult for regulators to assess whether the company is complying with its own stated principles. Some experts have called for mandatory third-party audits of AI safety processes, arguing that self-regulation has failed.
“We cannot continue to rely on the goodwill of a handful of private companies,” said Dr. Mei Lin, a technology policy researcher. “The disbandment of a team that was specifically created to worry about existential risks should be a political issue, not just a corporate announcement.”
What comes next
OpenAI is expected to continue its rapid rollout of new AI models and features in the coming months. The company has not officially confirmed the dismantling of the preparedness team, and representatives have declined to comment on the FT report. If the report is accurate, the change will take effect quietly, with no major press release or public statement—a notable distinction from the more transparent safety update the company used to publish on a regular basis.
For researchers and observers, the immediate concern is whether any warnings raised by the former preparedness team can still make their way to the top decision-makers. In the past, the team had a direct line to senior management and could escalate urgent issues. Under the new structure, that line becomes less direct, and the teams responsible for bio, cyber, and other domains will have to compete with the daily pressures of product development.
The broader trajectory is clear: OpenAI is evolving from a mission-driven research laboratory into a massive commercial enterprise. With each reorganization, the remaining safety infrastructure becomes more integrated into the business—and, critics would argue, more subservient to it. Whether the company has struck the right balance between innovation and caution is a question that will only be answered with time, and possibly with the consequences of the next major AI incident.
Source:The Verge News