The headline promises safety; the structure reveals decay. On October 2025, OpenAI dissolved its Preparedness team, the unit responsible for identifying and mitigating catastrophic risks from frontier models. This is not a cost-cutting footnote. It is the second major safety function to be dismantled in under twelve months, following the Superalignment team’s dissolution earlier in the year. The timing—ahead of an anticipated IPO—suggests a deliberate institutional recalibration. In my years auditing blockchain protocols, I have learned that when a project removes its independent security review function before a liquidity event, the codebase is not stronger; the trust layer is thinner. The same principle applies here.
Context: The Architecture of Safety at OpenAI
The Preparedness team was established in 2023 as a direct response to the growing recognition that frontier models pose existential risks—biological, cyber, persuasion, and autonomy. The team reported to the Safety and Security Committee under the board, acting as an internal adversarial check on model releases. Its dissolution means that the last dedicated independent unit for catastrophic risk assessment is gone. The Superalignment team was already absorbed into product teams, and now Preparedness follows. The narrative that these functions are “integrated” into broader development is precisely the kind of organizational camouflage that delays accountability. In blockchain, when a protocol merges its audit team into the core development channel, we call it a conflict of interest. Here, it is called restructuring.
Core: The Forensic Code of Organizational Governance
Let me dissect the structural implications systematically. First, loss of independent adversarial testing. The Preparedness team conducted red-team exercises that were not subject to product release pressure. Their evaluations could delay or block a launch if the risk threshold was exceeded. Without this buffer, the decision to ship a model becomes a business decision, not a safety decision. The differential between the two is where catastrophic failures incubate.
Second, talent flow. The team’s lead, Aleksander Madry, was already sidelined in late 2024. Key safety researchers have left for Anthropic, a competitor that has built its brand on safety-first positioning. The loss of institutional knowledge is not just a hiring cost; it is a knowledge gap in the corporate memory of how to assess emergent risks. In my experience, when a blockchain project loses its lead auditor, the next exploit is not a matter of if, but of when. The same temporal logic applies to AI models.
Third, the IPO signal. The reorganization is explicitly framed as preparatory for public listing. In capital markets, governance is a priced risk factor. The dissolution of a safety team signals to investors that the board prioritizes agility and cost savings over risk mitigation. This is a rational choice if the management believes the probability of a catastrophic safety event is low. But the history of AI—and of technology more broadly—is a history of underestimating tail risks. The Preparedness team was explicitly designed to correct that bias. Removing it is a structural bet that the bias is no longer dangerous.
Contrarian: What the Bulls Get Right
It is possible that the Preparedness team was ineffective. I have seen audit teams in crypto that produced checklists but no real security improvements. It is also possible that OpenAI is shifting safety evaluation to a more distributed model—embedding safety engineers into product teams, or outsourcing red-teaming to external firms. In that scenario, the reduction in headcount does not imply a reduction in safety rigor. Furthermore, the IPO pressure may force OpenAI to adopt external auditing standards that are more rigorous than internal ones. The EU AI Act, for example, requires third-party conformity assessments for high-risk models. If OpenAI is preparing for compliance, disbanding an internal team could be a step toward a more auditable, externalized safety framework.
But the structure of the decision reveals a deeper problem. In blockchain, when a DeFi protocol outsources its oracle risk to a single external provider, it gains flexibility but loses sovereignty. The same trade-off applies here. An external safety auditor is a vendor, not a fiduciary. The Preparedness team had a direct line to the board. An external firm has a contract. The difference is accountability. When the vendor is fired, the safety function disappears. When an internal team is dismantled, the safety function becomes a line item in a budget. The latter is what we are seeing.
Takeaway: The Hash of Governance
Structure reveals what emotion conceals. The dissolution of the Preparedness team is not a decision about safety versus speed. It is a decision about the concentration of safety decision-making. By removing the independent unit, OpenAI centralizes the authority to assess risk in the hands of product and business leaders. In a decentralized system, security is distributed. In a centralized one, it is a single point of failure. Investors should demand proof that the safety governance structure is not a facade. The blockchain remembers what you forget. The IPO filing will remember this decision.
Truth is found in the hash, not the headline. The headline reads “restructuring.” The hash reveals a systematic reduction in the institutional capacity to say no. In the next twelve months, we will see whether the market prices this risk correctly. I suspect the discount will come after the first unmitigated failure, not before.
Based on my audit experience, I have seen this pattern repeat across protocols. The team that controls the release decision becomes the team that downplays the risk. The Preparedness team was a check on that dynamic. Its removal is a structural vulnerability that will be exploited—not by hackers, but by time.