OpenAI has dissolved its Preparedness team, the group responsible for assessing whether increasingly capable AI models could create severe or catastrophic risks, according to a Financial Times report cited by The Decoder. The team was reportedly shut down at the end of July, with responsibility for areas including biological and cybersecurity risks redistributed to existing teams across the company.
The move marks another significant change in OpenAI’s approach to AI safety and comes at a sensitive moment for the company. OpenAI is undergoing a broad organizational restructuring as it prepares for a potential initial public offering, while its AI systems are becoming increasingly capable and autonomous. The company continues to maintain its Preparedness Framework and other safety mechanisms, but the dissolution of a dedicated team has raised questions about whether specialized oversight is being weakened or simply integrated more deeply into product and research organizations.
OpenAI Shuts Down Its Preparedness Team
The Preparedness team was created to identify, track and prepare for catastrophic risks associated with frontier AI models.
OpenAI reportedly dissolved the organizational unit at the end of July 2026.
Rather than eliminating the work altogether, the company has distributed specific responsibilities to existing teams.
| Key Detail | Information |
|---|---|
| Company | OpenAI |
| Team | Preparedness |
| Reported change | Team dissolved |
| Timing | End of July 2026 |
| Main role | Assess and mitigate catastrophic AI risks |
| Key risk areas | Cybersecurity, biological and other severe risks |
| New structure | Responsibilities distributed among existing teams |
| Former head | Dylan Scandinaro |
| Broader context | OpenAI organizational restructuring |
The change was reported by the Financial Times and subsequently covered by other technology publications.
What Was the Preparedness Team?
OpenAI established the Preparedness team in 2023 as a dedicated group focused on risks associated with increasingly capable AI systems.
Its original mandate included tracking, evaluating and forecasting dangerous capabilities that could emerge as frontier models became more powerful.
OpenAI identified areas including cybersecurity, chemical and biological threats, persuasion and autonomous replication and adaptation as potential risk categories.
Preparedness Mission
Monitor AI capabilities
↓
Identify dangerous capabilities
↓
Evaluate potential misuse
↓
Forecast future risks
↓
Develop safeguards
↓
Inform deployment decisions
The team was intended to provide specialized expertise around risks that could become significantly more severe as AI capabilities advanced.
OpenAI’s Original Preparedness Framework
When OpenAI launched the team, it described preparedness as a way to ensure that the company had the systems and infrastructure needed to manage catastrophic risks from frontier AI.
The framework was designed to connect capability evaluations, red-teaming, monitoring and mitigation.
OpenAI later updated the framework to establish more specific thresholds for high and critical capabilities. (openai.com)
The framework remains an important part of OpenAI’s stated safety approach.
The Work Is Being Redistributed
The reported restructuring does not mean that OpenAI has publicly said it is abandoning preparedness work.
Instead, responsibilities previously handled by the dedicated team are being assigned to existing groups.
Biological and cybersecurity preparedness work, for example, is reportedly being moved into specialized teams.
New Organizational Approach
Preparedness team
↓
Dissolved
↓
Specialized responsibilities
↓
Cybersecurity team
+
Biosecurity-related teams
+
Other existing groups
↓
Safety work continues
The key difference is that the work is no longer concentrated within one standalone organizational unit.
Why the Change Matters
The structure of an AI safety organization can influence how risks are identified and communicated.
A dedicated team can develop specialized expertise and focus exclusively on identifying dangerous capabilities.
When those responsibilities are distributed across existing teams, safety work can become more closely integrated with model development.
That could produce benefits, but it can also raise questions about independence and accountability.
Integration Could Make Safety Faster
There is a potential argument in favor of the restructuring.
If safety researchers work directly alongside teams building AI systems, potential risks can be identified earlier in the development process.
Instead of evaluating a model only after it has been developed, safety experts can participate throughout its creation.
Integrated Safety Model
Model development
↓
Safety evaluation
↓
Risk identified
↓
Engineering team responds
↓
New evaluation
↓
Improved model
This could shorten the feedback loop between capability development and risk mitigation.
Independent Oversight Could Become More Difficult
The opposite concern is that independent safety teams can provide a stronger challenge to product and research teams.
A standalone group may be better positioned to question whether a model should be released or whether additional safeguards are necessary.
If safety responsibilities are embedded inside development organizations, critics may worry that commercial or product priorities could influence risk assessments.
Independent Model
Product team
↓
Develops model
+
Independent safety team
↓
Tests model
↓
Challenges risk assumptions
↓
Leadership decision
Integrated Model
Product team
+
Safety specialists
↓
Develop and evaluate together
↓
Leadership decision
The difference is primarily organizational, but it can affect how disagreements are handled.
OpenAI Still Has a Preparedness Framework
It is important to distinguish the dissolution of the team from the disappearance of OpenAI’s preparedness framework.
In May 2026, OpenAI said its Preparedness Framework remained the foundation of its approach to managing serious risks from advanced AI systems.
The framework covers risks including cyber offense, chemical, biological, radiological and nuclear threats, harmful manipulation and loss of control. (openai.com)
Therefore, the organizational change does not by itself demonstrate that OpenAI has abandoned those safety commitments.
The Safety Advisory Group Still Has a Role
OpenAI’s framework includes a Safety Advisory Group, or SAG, which reviews risk assessments and safeguards for systems covered by the framework.
The group can make recommendations to OpenAI leadership regarding whether safeguards are sufficient and whether a model should be deployed.
The framework also gives the company’s leadership responsibility for final deployment decisions. (openai.com)
Existing Governance Structure
Preparedness evaluations
↓
Safety Advisory Group
↓
Risk assessment
↓
Safeguard recommendations
↓
OpenAI leadership
↓
Deployment decision
The question now is how the work previously performed by the dedicated Preparedness team will feed into this structure.
Dylan Scandinaro Is Moving to Other Research
Dylan Scandinaro, who joined OpenAI from Anthropic earlier in 2026 to lead Preparedness, is reportedly shifting his focus to research involving recursive self-improvement.
Scandinaro’s appointment had itself been viewed as a significant investment in frontier AI safety.
OpenAI announced his arrival in February, with Sam Altman saying he would help ensure the company’s systems were developed and deployed safely while preparing for risks from increasingly powerful AI. (indiatoday.in)
The Timing Is Significant
The restructuring comes as AI models are becoming increasingly capable of performing long-running and autonomous tasks.
OpenAI recently published research describing safety lessons from internal use of long-horizon models.
The company said these systems can solve difficult open-ended problems but also have more opportunities to take unwanted actions because they operate for longer periods. (openai.com)
This makes the timing of the Preparedness restructuring particularly notable.
Autonomous AI Creates New Safety Challenges
Earlier AI systems primarily generated text, images or code in response to individual prompts.
Newer systems can increasingly plan, use tools and perform multi-step tasks.
This creates a different category of safety risk.
Traditional AI Interaction
User
↓
Prompt
↓
AI response
↓
Interaction ends
Agentic AI
User
↓
Goal
↓
AI plans
↓
AI uses tools
↓
AI executes multiple steps
↓
AI observes results
↓
AI adapts
↓
Long-running activity
The longer an AI system operates autonomously, the more opportunities it has to make unexpected decisions.
OpenAI Has Recently Highlighted These Risks
OpenAI said in July that internal testing of a long-running model revealed novel failures that were not captured by its existing pre-deployment evaluations.
The company said it paused access and used the experience to develop new evaluations, improve long-horizon alignment and introduce trajectory-level monitoring. (openai.com)
This demonstrates that the company continues to invest in safety research even as its organizational structure changes.
Cybersecurity Is One of the Biggest Concerns
AI systems capable of generating and executing sophisticated code could potentially be misused for cyberattacks.
As models become better at programming, they could assist legitimate security researchers but also lower the barrier for malicious actors.
Preparedness work has therefore included evaluating the cybersecurity capabilities of frontier models.
AI Cybersecurity Risk
More capable model
↓
Better coding ability
↓
Greater automation
↓
Potential offensive cyber capability
↓
Higher potential impact
OpenAI has identified cybersecurity as one of the areas requiring preparedness measures.
Biological Risks Are Another Focus
Advanced AI systems can potentially assist researchers with scientific and biological tasks.
The same capabilities could also be misused.
For this reason, OpenAI’s Preparedness Framework has included biological and chemical risks among the categories requiring assessment.
The company has continued to publish safety work related to biological risks and has operated programs such as its Bio Bug Bounty.
Loss of Control Is a Different Category of Risk
Some AI safety researchers are concerned about systems becoming capable of taking actions that humans cannot easily predict or control.
OpenAI’s updated framework includes loss of control among the serious risk categories it considers.
This becomes increasingly relevant as models gain longer operating horizons and greater ability to interact with external systems.
OpenAI Has Already Restructured Safety Teams Before
The dissolution of Preparedness is not the first major change to OpenAI’s safety organization.
The company previously disbanded its Superalignment team in 2024 following the departure of co-leaders Ilya Sutskever and Jan Leike.
Other safety and alignment functions have also been reorganized.
These repeated changes have contributed to concerns among some former employees and outside observers about the company’s long-term safety priorities.
Former Employees Have Raised Concerns
Former OpenAI safety researcher Jan Leike has criticized the company’s direction, arguing that safety concerns could be receiving less attention as the company focuses on product development.
Other former employees have also raised concerns about changes to OpenAI’s safety culture.
However, these criticisms represent the views of former employees and do not establish that OpenAI has abandoned its safety work.
OpenAI Says It Is Moving Quickly
OpenAI has described its broader organizational changes as part of an effort to operate at an exceptional pace.
The company is competing aggressively in AI products, enterprise services and frontier model development.
As competition intensifies, organizational efficiency has become increasingly important.
Competitive Pressure
OpenAI
+
Anthropic
+
+
Meta
+
Other AI labs
↓
Rapid model development
↓
Shorter product cycles
↓
Need for efficient organizations
The challenge is maintaining rigorous safety processes while accelerating development.
IPO Preparations Add Another Layer
The restructuring also comes amid reports that OpenAI is preparing for a potential major public listing.
The Financial Times has reported that the company is undergoing multiple organizational changes as CEO Sam Altman prepares the business for a possible IPO.
This has led to speculation that OpenAI is trying to simplify its organizational structure and focus more heavily on core commercial priorities.
The company has not said that the Preparedness restructuring was made specifically because of an IPO.
Commercial Pressure Could Influence AI Safety
The AI industry is entering a phase in which frontier models are becoming major commercial products.
Companies face pressure to release increasingly capable systems quickly.
At the same time, regulators and the public are demanding stronger safety guarantees.
The Central Tension
Faster development
↓
More capable AI
↓
More commercial opportunities
BUT
↓
More potential risks
↓
Need for stronger safety
The organizational structure OpenAI chooses will determine how these competing priorities are managed.
Dedicated Teams vs Distributed Safety
There are two broad models for organizing AI safety.
A dedicated team provides centralized expertise and potentially greater independence.
A distributed approach embeds safety specialists directly within development teams.
Neither model is automatically safer.
The effectiveness depends on authority, resources, technical expertise and whether safety recommendations can actually influence deployment decisions.
The Real Question Is Authority
The most important issue may not be whether a standalone Preparedness team exists.
It is whether the people responsible for identifying catastrophic risks have enough authority to influence decisions.
A safety function can exist on paper but have limited impact if its recommendations can easily be overridden.
Effective Safety Governance
Risk identification
↓
Independent evaluation
↓
Clear escalation process
↓
Leadership review
↓
Ability to delay or modify deployment
↓
Accountability
The strength of this chain will matter more than the name or structure of a particular team.
OpenAI’s Board Safety Committee Remains Relevant
OpenAI’s Preparedness Framework also describes oversight by the Board’s Safety and Security Committee.
The committee can review major safety decisions and receive information from leadership.
The framework says the board can, when necessary, reverse a decision or require a revised course of action. (openai.com)
This provides an additional layer of governance beyond individual research teams.
AI Safety Is Becoming More Operational
The evolution of OpenAI’s safety structure reflects a broader industry shift.
AI safety is increasingly moving from a purely research-oriented discipline toward operational processes involving testing, monitoring, security controls and deployment decisions.
Modern AI Safety
Research
+
Red teaming
+
Capability evaluations
+
Monitoring
+
Security
+
Incident response
+
Governance
↓
Deployment safety
This broader approach could explain why some functions are being distributed across different organizational groups.
More Capable Agents Require Continuous Monitoring
Pre-deployment testing alone may not be sufficient for highly autonomous systems.
OpenAI itself has acknowledged that some behaviors can emerge during longer-running interactions that are not captured by fixed evaluations.
This makes runtime monitoring increasingly important.
Safety Lifecycle
Train model
↓
Evaluate model
↓
Deploy carefully
↓
Monitor behavior
↓
Detect unexpected behavior
↓
Update safeguards
↓
Re-evaluate
Safety therefore becomes an ongoing process rather than a one-time certification.
What It Means for OpenAI
For OpenAI, the restructuring could make safety work more integrated with product development.
It could also reduce organizational duplication and make teams more accountable for specific risk areas.
However, the company will need to demonstrate that the new structure maintains the expertise and independence required to identify severe risks.
What It Means for AI Safety Researchers
The move could influence how other AI companies organize safety teams.
If distributed safety proves effective, other companies may adopt similar models.
If critics are proven right that independence is essential, dedicated safety teams could become more important elsewhere.
What It Means for Regulators
Regulators are increasingly interested in how frontier AI companies assess and mitigate high-impact risks.
Organizational changes could therefore attract additional scrutiny.
Companies may need to demonstrate not just that they have safety frameworks, but that those frameworks are actually implemented and adequately resourced.
What It Means for Investors
Investors may view organizational streamlining positively if it reduces bureaucracy and accelerates product development.
However, significant reductions in independent safety capacity could create reputational, regulatory and operational risks if a major AI incident occurs.
The balance between speed and safety will therefore remain an important consideration for the company’s long-term valuation.
What Investors and Industry Should Watch
Key developments include:
- How OpenAI redistributes Preparedness responsibilities
- Whether the Preparedness Framework changes
- The role of the Safety Advisory Group
- Board Safety and Security Committee oversight
- Cybersecurity evaluations
- Biological-risk evaluations
- Long-horizon model testing
- Runtime monitoring
- Safety staffing levels
- Further executive and organizational changes
- OpenAI’s potential IPO plans
The effectiveness of the new structure will ultimately be judged by how well it identifies and mitigates dangerous capabilities.
Key Facts at a Glance
| Issue | Detail |
|---|---|
| Team affected | Preparedness |
| Reported dissolution | End of July 2026 |
| Created | 2023 |
| Original mission | Track and mitigate catastrophic AI risks |
| Major risk areas | Cyber, biological, CBRN, autonomy and loss of control |
| Former head | Dylan Scandinaro |
| New approach | Responsibilities distributed among existing teams |
| Preparedness Framework | Still maintained by OpenAI |
| Safety Advisory Group | Remains part of governance framework |
| Board oversight | Safety and Security Committee |
| Broader context | Organizational restructuring and potential IPO |
Infographic: OpenAI’s Safety Restructuring
2023
↓
OPENAI CREATES
PREPAREDNESS TEAM
↓
MISSION
↓
TRACK FRONTIER AI CAPABILITIES
+
EVALUATE CATASTROPHIC RISKS
+
DEVELOP SAFEGUARDS
↓
KEY AREAS
↓
CYBERSECURITY
+
BIOLOGICAL / CBRN
+
AUTONOMY
+
LOSS OF CONTROL
↓
2026
↓
PREPAREDNESS TEAM DISSOLVED
↓
RESPONSIBILITIES REDISTRIBUTED
↓
EXISTING SPECIALIZED TEAMS
↓
BUT
↓
PREPAREDNESS FRAMEWORK
REMAINS
↓
SAFETY ADVISORY GROUP
+
BOARD SAFETY & SECURITY COMMITTEE
↓
CONTINUED SAFETY OVERSIGHT
The Bigger Picture
OpenAI’s decision to dissolve its dedicated Preparedness team represents a significant organizational change in how one of the world’s leading AI companies handles catastrophic-risk research. The team was created in 2023 to track dangerous capabilities in frontier models and develop safeguards against risks including cyberattacks, biological threats and loss of control. According to reporting from the Financial Times, its work is now being redistributed to existing teams rather than eliminated altogether. :contentReference[oaicite:0]{index=0}
The move does not mean OpenAI has abandoned its formal safety framework. The company said as recently as May 2026 that its Preparedness Framework remains the foundation for managing serious risks from advanced AI, while its Safety Advisory Group and board-level Safety and Security Committee remain part of the governance structure. :contentReference[oaicite:1]{index=1} The more important question is therefore whether distributed safety responsibilities can preserve the expertise, independence and authority that a dedicated team provided. That question becomes increasingly important as AI systems gain longer operating horizons, stronger coding abilities and greater autonomy.
Looking Ahead
OpenAI’s next challenge will be proving that the new structure can maintain rigorous safety evaluations while allowing research and product teams to move quickly. The company has recently acknowledged that long-running AI systems can produce unexpected behaviors that existing pre-deployment evaluations may miss, reinforcing the need for continuous testing, monitoring and intervention. :contentReference[oaicite:2]{index=2}
Over the longer term, the restructuring could become a model for how frontier AI companies organize safety work. If risk specialists embedded within development teams can identify problems earlier and coordinate safeguards more effectively, the distributed approach could become increasingly common. But if organizational integration reduces independent scrutiny or makes it harder for safety researchers to challenge commercial decisions, pressure for stronger external oversight could increase. As AI capabilities continue advancing, the effectiveness of OpenAI’s safety governance will ultimately depend less on whether it has a team called Preparedness and more on whether its safety processes have the authority and resources to stop or modify unsafe deployments.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
