Hook
On a quiet Tuesday afternoon, the news hit the wires: OpenAI, the company that has become synonymous with frontier AI, disbanded its Preparedness team. This is the team tasked with assessing catastrophic risks—bioterrorism, cyberattacks, persuasion at scale, autonomous replication. The same team that, just a year ago, was presented as the crown jewel of the company's safety architecture. Now, it's gone.
I have seen this pattern before. In 2017, I watched a crypto fund gut its risk analysis team two months before a major ICO. The team was told their work was being "integrated into core operations." What followed was a 90% capital loss. The code did not lie, but the contract did. And when governance structures are sacrificed for a clean balance sheet, the rot is always beneath the yield.
OpenAI is not a blockchain project. But the engineering of trust is identical. When a company that sells safety as a competitive advantage quietly removes the team that ensured that safety, the market should listen. Not to the press release, but to the silence that follows.
Context
OpenAI's Preparedness team was established in late 2023, led by Aleksander Madry. Its mandate was to identify, evaluate, and mitigate "catastrophic risks" from frontier models. It reported directly to the board-level Safety and Security Committee. This was not a peripheral unit; it was the central nervous system of OpenAI's risk governance. The team's dissolution, announced as part of a broader restructuring ahead of an anticipated IPO, is the second major safety function contraction in less than a year. The Superalignment team was similarly dispersed in early 2024.
The IPO preparation is the key context. OpenAI has been transitioning from a capped-profit non-profit to a for-profit benefit corporation (PBC). This requires a leaner cost structure, a clearer growth narrative, and a balance sheet that satisfies institutional investors. Security teams, especially those with no direct revenue generation, are natural targets for cost optimization. The question is not whether the team was dissolved, but whether the function was replaced.
According to the article that triggered this analysis, the author suggests that the restructuring may dilute AI safety focus and raise public concerns about the company's risk management. That is a reasonable inference. But the article does not answer the critical question: did the safety function move, or did it evaporate?
Core: Systematic Teardown of the Dissolution
Let me be clear: the dissolution of a risk team is never a neutral event. In my years auditing blockchain protocols, I have learned that when a company disbands a compliance or security unit, it is usually because the unit's recommendations were slowing down product launches. The Preparedness team's job was to say "no" or "not yet." In a pre-IPO environment, "not yet" is a luxury the company can no longer afford.
First, the organizational signal. The Preparedness team was a direct line to the board. By removing that line, OpenAI is centralizing risk decisions in the executive team—the same team that is incentivized by revenue growth and market share. This is a classic principal-agent problem. The board's oversight of catastrophic risk is now mediated through the same people who are evaluated on quarterly earnings. The code does not lie, but the governance structure can.
Second, the talent impact. Ilya Sutskever, Jan Leike, and now the Preparedness team's rank and file. The pattern of safety researchers leaving OpenAI is not a coincidence. Leike went to Anthropic. Others will follow. The safety community is small; knowledge is tacit. When these people leave, they take with them the institutional memory of how to evaluate frontier risks. You cannot replace that with a contract worker or an external auditor who sees the model for a week. Silence is the loudest indicator of risk.
Third, the compliance vacuum. The Preparedness team was responsible for producing the "System Card" risk assessments that accompany each major model release. These documents are not just PR; they are used by enterprise customers—banks, hospitals, government agencies—to conduct their own due diligence. If the team is gone, who writes the next System Card? Will it be a self-assessment by the product team? Will it be outsourced to a third party that lacks the engineering depth to understand the model's internals? The article does not say, but the absence of this information is itself a data point.
Fourth, the timing. The IPO is reportedly expected in 2025 or 2026. The Preparedness team's dissolution is happening now, well before the S-1 filing. This is strategic. The company wants to show a clean organizational chart and a streamlined cost structure to investors. Safety teams are expensive—they require top-tier researchers, massive compute budgets for red teaming, and the authority to delay releases. By removing the team now, OpenAI can avoid uncomfortable questions from underwriters about potential launch delays due to safety reviews.
Fifth, the narrative shift. OpenAI has built its brand on being the "safe" AI company. Sam Altman's testimony before Congress, the charter of the company, the public commitment to responsible deployment—all of that was a narrative that required a visible safety team. By dissolving the team, OpenAI is signaling that the narrative is no longer the priority. The product is the priority. Hype is noise; structure is signal. The structure here says: revenue first, safety second.
Contrarian Angle: What the Bulls Got Right
But I am not here to write a one-sided panic. The bulls have a point. First, the Preparedness team's dissolution does not necessarily mean the end of safety work. The function may be absorbed into a new "Model Release Board" or a centralized deployment team. In large organizations, restructuring can sometimes improve efficiency by removing duplication. The Preparedness team may have been redundant with other groups, such as the policy team or the engineering safety group.
Second, external safety assessment is becoming a market. If OpenAI outsources its red teaming to independent firms, it could actually increase the rigor of evaluations by introducing fresh eyes. The blockchain world has seen this: internal auditors often miss blind spots that external auditors catch. The key is whether the outsourced evaluators have the same level of access to the model's internals and the same authority to block a release.
Third, the IPO pressure may force OpenAI to adopt more formal governance standards, not fewer. Public companies face stricter requirements for risk management, board oversight, and disclosure. The Preparedness team's dissolution could be a prelude to a more robust, publicly accountable safety framework that includes independent board members with technical expertise. The current structure was opaque; a public company structure may be more transparent.
Fourth, the models themselves are becoming more aligned through reinforcement learning from human feedback (RLHF) and constitutional AI. The safety research that the Preparedness team was doing may have already been incorporated into the training pipeline. The team's work may be complete, or at least at a stage where it can be maintained by a smaller group. Beauty is the mask; geometry is the bone. The geometry of alignment may be solid enough to withstand the loss of a single team.
Takeaway: The Accountability Call
I do not follow the wave; I measure its depth. The depth here is troubling. The dissolution of the Preparedness team is not a minor reorganization; it is a governance choice that prioritizes short-term financial optics over long-term risk management. The blockchain industry learned this lesson the hard way: teams that cut safety to chase growth end up with hacks, losses, and regulatory crackdowns. AI is no different.
If OpenAI is serious about safety, it will announce a credible replacement structure within 90 days. If not, the market should treat this as a red flag. The code does not lie, but the governance can. And when governance fails, the rot is already beneath the yield.
I will be watching the next System Card. If it is shorter, or if it is absent, the silence will speak volumes.