OpenAI has shut down its Preparedness team, the group responsible for assessing severe and potentially catastrophic risks associated with increasingly capable artificial intelligence models.
The organizational change, reported in late July 2026, moves responsibilities such as biological and cybersecurity preparedness into existing teams instead of keeping them under one dedicated Preparedness organization.
The decision comes at an important moment for OpenAI. AI models are becoming more capable at coding, scientific research, reasoning, tool use and autonomous tasks. The rise of more capable AI agents that can perform complex workflows makes those capabilities increasingly relevant to safety discussions.
At the same time, the company is undergoing broader organizational changes and facing growing pressure to expand commercial products and enterprise services.
That combination has made the OpenAI Preparedness Team Shutdown an important development for the AI safety community.
The shutdown does not automatically mean that OpenAI has stopped working on AI safety. Instead, the company has changed how preparedness responsibilities are organized.
The bigger question is whether distributing those responsibilities across existing teams will make safety more deeply integrated into AI development or make accountability and independent risk assessment more difficult.
What Was OpenAI’s Preparedness Team?
OpenAI’s Preparedness team was created to focus on severe risks associated with advanced AI models.
Unlike conventional product safety work, Preparedness focused on scenarios in which increasingly capable models could potentially create large-scale or catastrophic harm.
Its responsibilities included evaluating dangerous capabilities, identifying emerging risks and developing strategies to mitigate those risks before advanced systems were widely deployed.
The team worked across several major areas, including biological risks, cybersecurity threats and other frontier AI risks.
OpenAI’s broader safety framework provides useful context for understanding this work. The company’s current Frontier Governance Framework says its Preparedness Framework remains the foundation for managing serious risks from advanced AI systems.
The framework covers areas including cyber offense, CBRN risks, harmful manipulation and loss of control.
That means the shutdown of a specific organizational unit should not be confused with the disappearance of the underlying safety framework.
The distinction between a team and a safety function is central to understanding what happened.
What Did the Preparedness Team Work On?
The Preparedness team covered several categories of frontier AI risk.
Biological and Bioweapon Risks
One major area was biological risk.
As AI systems become better at scientific reasoning, research assistance and technical problem-solving, safety researchers need to determine whether these capabilities could be misused to increase biological threats.
The goal of preparedness research was not simply to assume that an AI model would cause harm. Instead, researchers evaluated what capabilities a model actually demonstrated and considered whether those capabilities could meaningfully increase existing risks.
This type of assessment becomes more important as AI moves beyond basic text generation and begins supporting more complex AI-powered scientific research workflows.
As scientific AI agents become more capable, researchers must consider not only what these systems can accomplish but also what safeguards may be necessary around increasingly autonomous scientific work.
Cybersecurity and Cyberattack Risks
Cybersecurity was another important part of Preparedness work.
Modern AI models can write and analyze code, identify vulnerabilities, interact with tools and perform increasingly complex multi-step tasks.
These capabilities can benefit legitimate cybersecurity teams, but they can also create new opportunities for malicious use.
The growing use of AI agents for vulnerability discovery and cybersecurity shows how quickly AI is becoming integrated into security workflows.
Preparedness research therefore considered whether advanced AI systems could assist sophisticated cyber operations or eventually perform parts of those operations with greater autonomy.
OpenAI continues to treat cybersecurity as a major frontier risk. Its recent safety documentation describes cybersecurity evaluations and safeguards as part of its deployment process for highly capable models.
Loss-of-Control Risks
Preparedness also examined broader questions surrounding loss of control.
This involves considering hypothetical situations in which increasingly capable AI systems could become difficult for humans to reliably control, constrain or monitor.
These scenarios remain an area of active research rather than established predictions about current AI systems.
The purpose of evaluating them is to understand potential failure modes before models become more capable.
OpenAI’s current safety approach similarly identifies loss of control and autonomy as important areas of frontier risk assessment.
Automated Red Teaming
Another important component of frontier AI safety is red teaming.
Red teaming involves deliberately testing AI systems to identify weaknesses, dangerous capabilities and unexpected behaviors before deployment.
Automated red teaming can expand this process by allowing researchers to test models at much greater scale.
Instead of relying exclusively on manual testing, researchers can use automated systems to explore large numbers of potential failure cases.
OpenAI’s Preparedness Framework has continued to emphasize ongoing evaluations and red teaming as AI capabilities evolve.
Why Did OpenAI Shut Down the Preparedness Team?
OpenAI has not characterized the change as abandoning preparedness or AI safety.
Instead, the company has indicated that safety and security responsibilities are being integrated more deeply into model development.
According to reporting cited in the original article, the Preparedness team was disbanded at the end of July, with senior staff taking responsibility for specific areas such as biological and cyber preparedness within existing teams.
Greg Brockman also described the broader organizational direction as integrating research, safety and security more deeply into model development.
That distinction is important.
The organizational structure changed; the underlying risk-assessment work did not simply disappear.
A distributed safety model can have practical advantages.
If safety specialists work directly with engineers and researchers building advanced models, potential problems can potentially be identified earlier in the development process.
Instead of treating safety as a separate checkpoint, the company can make risk assessment part of everyday model development.
This broader shift also reflects how AI safety is becoming increasingly multidisciplinary, bringing together expertise from machine learning, mathematics, cybersecurity, policy and other fields.
However, this structure also creates questions.
A dedicated safety organization has an obvious mandate, specialized expertise and clearly identifiable leadership.
When responsibilities are distributed among multiple teams, it becomes more important to understand who owns each risk category and whether those researchers have enough authority to challenge product or deployment decisions.
What Happened to the Preparedness Team’s Work?
The Preparedness team’s work was redistributed rather than simply discontinued.
Responsibilities involving areas such as biological and cyber preparedness were moved into existing parts of OpenAI.
This approach is consistent with OpenAI’s broader stated philosophy that safety should be integrated throughout the development and deployment process.
OpenAI says its safety approach involves continuous evaluations during training and deployment rather than relying on one safety review at the end of development.
There are potential benefits to this model.
Safety researchers who work closely with model-development teams may have better access to technical information.
They may also be able to influence decisions earlier, when changing a model or adding safeguards could be easier.
But integration alone does not guarantee effective safety.
The key issues are ownership, resources, expertise, independence and authority.
If every major risk area has a clearly responsible leader and sufficient technical resources, distributed preparedness could work effectively.
If responsibilities become fragmented across teams without clear accountability, important risks could become harder to track.
What Is Dylan Scandinaro Focusing on Now?
Dylan Scandinaro, who previously led OpenAI’s Preparedness effort, has shifted his focus toward the safety implications of recursively self-improving AI, according to reporting on the restructuring.
Scandinaro joined OpenAI as Head of Preparedness in February 2026 after previously working on AI safety at Anthropic.
His appointment was presented as part of OpenAI’s effort to strengthen its preparation for severe AI risks.
His new area of focus raises another important frontier AI question.
What Is Recursively Self-Improving AI?
In simple terms, recursively self-improving AI describes a hypothetical scenario in which an AI system contributes to improving its own capabilities or helps create more capable AI systems.
For example, an advanced system could potentially assist with algorithm optimization, training techniques, software development or the creation of another AI model.
If such processes became increasingly automated and were repeated multiple times, researchers would need to understand whether AI capabilities could improve faster than existing safety mechanisms could adapt.
This is a research and risk-assessment topic, not evidence that unrestricted recursive self-improvement is currently taking place.
Does the Shutdown Mean OpenAI Is Deprioritizing AI Safety?
Not necessarily.
The available evidence supports a more nuanced interpretation.
OpenAI continues to publish safety frameworks, system cards, risk evaluations and deployment assessments.
Its current safety materials state that the Preparedness Framework guides decisions around serious risks from advanced AI systems.
For example, OpenAI’s GPT-5.5 System Card says the model underwent pre-deployment safety evaluations under the Preparedness Framework, including targeted red teaming for advanced cybersecurity and biology capabilities.
That evidence suggests that preparedness work remains part of the model-development process.
At the same time, the organizational restructuring creates legitimate questions about independence and accountability.
There are two broad ways to interpret the change.
The first is safety integration.
OpenAI may believe safety works better when specialists are embedded directly into teams responsible for developing advanced AI systems.
The second is reduced centralization.
Moving responsibilities away from a dedicated organization could make it harder for outside observers to identify who is responsible for catastrophic-risk assessment and how independently those researchers can challenge decisions.
The most balanced conclusion is that OpenAI is changing how it organizes AI safety, while the long-term effect of that change remains uncertain.
Why the Preparedness Shutdown Matters Now
The timing makes the restructuring particularly significant.
AI systems are becoming more capable across the same areas that make frontier-risk evaluation important.
Models can now write sophisticated software, analyze technical information, interact with external tools and complete increasingly complex multi-step tasks.
These capabilities create substantial benefits, but they can also introduce new risks.
Cybersecurity is a clear example.
A model that becomes better at vulnerability discovery, code generation and autonomous task execution may become more useful to defenders while simultaneously becoming more valuable to attackers.
The emergence of tools such as AI systems designed to find exploitable software vulnerabilities illustrates how advanced models can increasingly participate directly in security work.
The same principle applies to scientific and biological capabilities.
The stronger models become, the more important it becomes to measure their capabilities before deployment rather than assuming that traditional safety testing will remain sufficient.
OpenAI’s current safety philosophy reflects this approach.
The company says it continuously expands and refines evaluation suites as capabilities and usage evolve.
What Does the Shutdown Reveal About OpenAI’s Priorities?
The Preparedness restructuring can be interpreted through three major themes.
1. Safety Integration
OpenAI may believe safety should be embedded directly into model-development workflows.
Under this approach, researchers evaluating cyber, biological or autonomy risks can work closer to the engineers developing the underlying models.
This could make safety testing more practical and responsive.
2. Organizational Streamlining
The shutdown also comes during a broader period of organizational change at OpenAI.
The Financial Times reported significant internal upheaval and restructuring as the company prepares for a potential IPO and expands its commercial operations.
The report also linked the preparedness restructuring to wider concerns about the company’s changing priorities.
That broader context makes it difficult to evaluate the Preparedness shutdown in isolation.
3. Greater Commercial Pressure
OpenAI is operating in an increasingly competitive AI market.
The company is expanding products, enterprise services and infrastructure while competing with other major AI developers.
The rapid development of advanced AI agents for workplace tasks demonstrates how quickly frontier models are moving from conversational tools toward systems capable of acting across software, documents and business workflows.
Commercial pressure does not automatically mean safety is being ignored.
However, as AI companies move faster, the challenge is ensuring that safety evaluation keeps pace with increasingly rapid model development.
The critical question is therefore not simply whether OpenAI maintains a team called Preparedness.
The more important question is whether the company continues to provide strong evaluations, clear accountability, independent challenge and sufficient resources for catastrophic-risk research.
What the Shutdown Means for Developers and AI Users
For developers and everyday AI users, the Preparedness shutdown does not mean OpenAI models suddenly became unsafe.
The more useful question is how OpenAI evaluates increasingly powerful models before and after deployment.
Developers and organizations using advanced AI should pay attention to:
- AI safety evaluations
- Model risk assessments
- Preparedness Framework updates
- Cybersecurity testing
- Biological and CBRN evaluations
- System cards and safety documentation
- Red-team results
- Deployment safeguards
- Changes in model autonomy and tool access
OpenAI’s published system cards provide one way to follow these developments.
For example, its o3 and o4-mini System Card describes Preparedness evaluations across biological and chemical capability, cybersecurity and AI self-improvement.
This type of documentation can provide a clearer picture of safety practices than the existence or absence of a particular organizational team.
Readers who want to understand how these developments fit into the broader market can also explore an AI tools and technology directory covering different platforms and applications.
What to Watch Next
The next stage will show whether OpenAI’s new structure delivers the level of preparedness the company says it wants.
Several developments will be particularly important.
First, readers should watch for updates to OpenAI’s Preparedness Framework and related governance documents.
Second, new model system cards may reveal how the company continues to evaluate cybersecurity, biological risks, autonomy and AI self-improvement.
Third, changes in leadership and responsibility will matter.
If risk areas receive clear owners with sufficient authority and resources, the distributed model may prove effective.
Finally, Dylan Scandinaro’s work on recursively self-improving AI could become an important indicator of how OpenAI is thinking about emerging frontier risks.
OpenAI’s latest governance framework makes clear that its approach is expected to evolve as model capabilities, evaluations and regulatory requirements change.
The wider debate over how AI safety frameworks should govern frontier models also shows that accountability, testing and risk thresholds are becoming important questions beyond any single AI company.
The Bigger Picture for AI Safety
The OpenAI Preparedness Team Shutdown is significant because it highlights a larger challenge facing the AI industry.
As models become more capable, safety organizations must decide how to balance independence with integration.
A completely separate safety team may have stronger organizational independence, but it can also be farther away from the engineers building the technology.
An embedded model can bring safety expertise closer to development, but it may create concerns about whether researchers have enough independence to stop or challenge a deployment.
Neither structure is automatically safer.
The effectiveness of the system depends on what authority researchers have, how risks are measured, how transparent the evaluations are and whether leadership acts on identified problems.
That is why the future performance of OpenAI’s new structure will matter more than its organizational name.
Conclusion
The OpenAI Preparedness Team Shutdown represents a major change in how OpenAI organizes work on catastrophic AI risks.
The company has not said that it is abandoning AI safety.
Instead, preparedness responsibilities have been distributed among existing teams, including areas such as biological and cybersecurity risk.
At the same time, former Preparedness leader Dylan Scandinaro has shifted his attention toward the safety implications of recursively self-improving AI.
The restructuring therefore should not be viewed as simple evidence that OpenAI has stopped caring about AI safety.
The more important test is what happens next.
If the new structure produces stronger evaluations, faster risk detection and closer cooperation between safety researchers and development teams, it could strengthen OpenAI’s approach to frontier AI safety.
If specialized expertise, clear ownership or independent oversight becomes weaker, concerns about the company’s safety priorities are likely to grow.
For now, the most accurate interpretation is that OpenAI is changing how it manages AI preparedness, not necessarily abandoning the work itself.
As AI systems become more powerful, rigorous testing, transparent risk assessment and strong safeguards will remain essential regardless of whether that work sits inside a dedicated Preparedness team or across multiple parts of the company.
