OpenAI has reportedly dismantled the dedicated team charged with anticipating the most serious dangers posed by increasingly capable artificial-intelligence systems, redistributing its work across existing specialist groups. The move alters the organizational machinery behind the company’s public commitments to evaluate frontier models for severe risks before they are released.

The Preparedness team was disbanded at the end of July, according to the Financial Times, whose reporting was subsequently summarized by The Verge on August 16. Responsibilities associated with particular threat areas, including biological and cybersecurity risks, were reassigned to teams already working in those domains. OpenAI had not published a detailed organizational chart explaining the new division of authority as of the reports.

The change is best understood as a restructuring of safety work rather than the announced abandonment of the underlying risk program. OpenAI’s Preparedness Framework remains a published company policy. Yet the separation between a formal framework and the people empowered to operate it is significant: rules governing dangerous capabilities depend on clear ownership, reliable escalation and the ability of evaluators to challenge research or product teams under pressure to advance a model.

Preparedness was designed to address risks that extend beyond familiar concerns such as inaccurate answers, biased outputs or routine abuse. Its mandate focused on frontier capabilities that could create severe or potentially irreversible harm. OpenAI’s current framework identifies biological and chemical capabilities, cybersecurity capabilities and AI self-improvement as established “Tracked Categories.” It also identifies developing research areas such as long-range autonomy, autonomous replication and adaptation, deliberate underperformance, attempts to undermine safeguards, and nuclear or radiological risks.

That distinction matters because the relevant systems are no longer merely conversational products. Frontier models can write and execute code, operate digital tools, assist with laboratory reasoning and complete multistep tasks with decreasing human supervision. Improvements in those areas create commercial value for software development, scientific research and enterprise automation. The same capabilities may also lower barriers for attackers, amplify specialist knowledge or allow agents to take consequential actions in complex environments.

OpenAI revised its Preparedness Framework in April 2025 to concentrate on risks it described as plausible, measurable, severe, newly enabled by advanced models, and potentially instantaneous or difficult to reverse. The framework streamlined capability classifications around two consequential thresholds. A system reaches “High” capability when it could amplify existing routes to severe harm. A “Critical” capability could introduce unprecedented pathways to such harm.

The commitments attached to those thresholds are more important than the labels. Under the published framework, a covered system assessed at High capability must have safeguards that sufficiently minimize the associated risk before deployment. A system reaching Critical capability is subject to a stricter standard: adequate safeguards are required during development as well as before external release. This approach recognizes that a sufficiently powerful model can create security concerns even while it remains inside a laboratory.

The framework assigns an important review function to OpenAI’s Safety Advisory Group, a cross-functional body of internal safety leaders. That group is meant to examine capabilities and safeguards reports, assess remaining risk and recommend whether a model can proceed, requires more evaluation or needs stronger protections. Final decisions remain with OpenAI’s leadership. The company has not publicly indicated that the Preparedness team’s dissolution eliminates the advisory group or changes those formal decision rights.

However, distributing evaluation work among domain teams may change how evidence reaches that review process. An embedded model can have advantages. Cybersecurity specialists working directly with researchers may detect new offensive capabilities earlier and help design technical controls more quickly. Biological-risk experts placed near scientific-model development may better understand how a capability emerged and which restrictions are practical. Integrating evaluators with engineering teams can also reduce delays caused by handoffs between organizational silos.

The countervailing risk is fragmentation. A centralized preparedness function can compare findings across domains, identify threats that do not fit cleanly into a single category and preserve institutional memory about earlier evaluations. It can also provide a recognizable point of accountability when a model exhibits several dangerous capabilities simultaneously. Once responsibility is divided, outsiders may find it harder to determine who owns the complete risk picture or who has authority to delay development.

Independence is another concern. Safety specialists embedded within product or research organizations may gain technical context, but they can also face the same deadlines, incentives and leadership structures as the teams whose systems they assess. That does not mean their conclusions will be compromised. It does mean governance must compensate with documented escalation procedures, review bodies capable of resisting schedule pressure and transparent evidence showing that safeguards work in practice.

Dylan Scandinaro, the reported head of the Preparedness team, is expected to shift his attention to the implications of recursively self-improving AI. Scandinaro joined OpenAI from Anthropic in February 2026, when Chief Executive Sam Altman publicly emphasized the urgency of preparing for severe risks from powerful models. His new focus remains closely connected to the framework’s AI self-improvement category, but it is narrower than the dissolved unit’s cross-domain remit.

Recursive self-improvement refers broadly to systems contributing to the research, engineering or automation that produces more capable successor systems. The concept ranges from AI assistance that speeds routine model development to more autonomous cycles in which systems meaningfully accelerate improvements to AI itself. Measuring that capability is difficult because progress may arise from combinations of coding ability, research judgment, tool use, long-horizon planning and access to computing infrastructure rather than from a single benchmark result.

The restructuring follows other changes to OpenAI’s safety-oriented organizations. The company disbanded its Superalignment team in 2024 after co-leaders Ilya Sutskever and Jan Leike departed. That initiative had been established to study how humans might control systems more intelligent than themselves. OpenAI also redistributed work previously handled by its AGI Readiness organization after adviser Miles Brundage left in 2024. A Mission Alignment function was later folded into other operations.

Those organizational changes have been accompanied by turnover among people associated with safety, ethics and mission governance. Recent reports have cited the departures of ethics lead Chloé Bakalar, chief futurist Joshua Achiam and safety-systems leader Johannes Heidecke. Departures can reflect varied personal and professional reasons, and they do not by themselves establish a retreat from safety spending. Taken together with repeated team restructurings, however, they have increased scrutiny of continuity and internal checks.

Former OpenAI researcher Jan Leike has been among the prominent critics of the company’s balance between safety work and product development. After leaving in 2024, he argued publicly that safety had taken a back seat to product priorities. OpenAI has maintained that integrating safety more deeply into research can improve decisions by ensuring that capability and risk assessments inform each other. The Preparedness reorganization places that argument under a new practical test.

The timing also gives the decision a business dimension. OpenAI is operating in a capital-intensive race with Anthropic, Google, Meta and other developers to produce more capable systems and secure enterprise adoption. Training and serving frontier models require vast expenditures on chips, data centers, energy, networking and specialist talent. The Financial Times reported that OpenAI’s broader reorganization is unfolding as the company prepares for a potential public listing, although the timing and terms of any offering remain uncertain.

Prospective public-market investors would assess rapid revenue growth alongside substantial infrastructure costs, competitive pressure, leadership stability and regulatory exposure. Safety governance increasingly intersects with each of those considerations. A serious model incident could disrupt releases, damage enterprise trust, invite litigation or produce binding regulatory requirements. Conversely, excessively slow or poorly coordinated internal review could impede commercialization in a market where model performance changes quickly.

Enterprise customers have their own reasons to examine the restructuring. Businesses using AI agents for coding, research or operational workflows rely on vendors to identify capabilities and failure modes that customers cannot independently test. A published policy is useful, but procurement and risk teams may seek clearer evidence regarding evaluation coverage, external testing, incident reporting and the authority of internal safety reviewers. The distribution of preparedness work could strengthen those processes if domain teams produce more technically rigorous controls, or weaken confidence if accountability becomes opaque.

Regulators are likely to focus less on the names of internal teams than on demonstrable outcomes. Relevant questions include whether OpenAI continues to test models before and during deployment, whether dangerous findings reach senior decision-makers, whether safeguards are independently challenged, and whether the company reports material changes in capability. The effectiveness of the new structure will therefore be judged by evaluation quality, enforcement and disclosure rather than headcount or organizational branding alone.

OpenAI’s framework includes Capabilities Reports intended to assess whether a model has crossed a defined threshold and Safeguards Reports describing protections and their effectiveness. The company said in 2025 that it would publish preparedness findings with frontier-model releases. Continued publication would provide one visible measure of whether the redistributed teams can maintain the framework’s operational cadence. The detail, consistency and timeliness of those reports will matter as much as their existence.

Another test will be how OpenAI handles risks spanning multiple categories. A highly autonomous coding agent, for example, could combine cyber capabilities with long-range planning, safeguard evasion and contributions to further AI development. Domain-specific teams may each assess part of that behavior, but leadership still needs an integrated view of the combined risk. Cross-functional governance becomes more—not less—important when technical responsibility is decentralized.

The debate should not be reduced to whether centralized or embedded safety teams are inherently superior. Either structure can fail without resources, authority and strong incentives. A standalone group may become isolated from core research or receive information too late. An embedded system may normalize risks or allow ownership to diffuse. The decisive factors are whether evaluators have access to models and data, whether adverse results can halt work, and whether unresolved disagreements are recorded and escalated.

OpenAI’s next major capability assessments will provide the clearest evidence. Observers will look for continuity in threshold testing, safeguards analysis and Safety Advisory Group review. They will also watch whether the company identifies accountable leaders for biological, cyber and self-improvement risks, and whether external evaluators receive sufficient access to verify important conclusions.

For now, the reported dissolution replaces one visible safety institution with a more distributed operating model. OpenAI says, through its published framework and previous statements, that preparedness remains integral to developing advanced AI. Critics see the loss of another dedicated unit as a warning that independent safety functions are being weakened as commercial pressures rise. The credibility of each position will depend on what the reorganized teams do when a model approaches a High or Critical threshold—and whether their authority is strong enough to change the company’s course.