The Safety Silo Collapse: Why OpenAI’s Organizational Surgery Signals a Deeper Rot
You are mistaken if you think this is just a personnel change. It’s a structural reconfiguration of OpenAI’s risk architecture. The decision to absorb the safety team into the research VP’s reporting line, coupled with the departures of Ilya Sutskever and Jan Leike, is not a minor reshuffle. It is a clear declaration that the company’s internal debate between safety-first idealism and commercial acceleration has been resolved. And the winners are the people who ship code, not the ones who audit it.
Let me walk you through the evidence. Over the past 18 months, I’ve traced the decay of OpenAI’s governance credibility through three distinct data points: the dissolution of the Superalignment team, the quiet removal of independent safety reporting lines, and the public resignations of the two most prominent figures associated with long-term AI risk. Each event is a block in a chain that leads to an inevitable conclusion: OpenAI has structurally subordinated safety to speed. And when that happens, the system becomes susceptible to the same kind of catastrophic failures that have plagued every over-leveraged protocol in crypto history.
Context: The Superalignment team was launched in July 2024 with a bold mandate — solve the alignment of superhuman AI within four years. It was co-led by Ilya Sutskever, the company’s chief scientist and a co-founder, and Jan Leike, a former DeepMind researcher who had built the RLHF pipeline from scratch. They had direct access to the board. They had independent budget. They represented the thesis that technical alignment was a non-negotiable prerequisite for AGI deployment. That thesis is now dead. The team was disbanded in early 2026. Leike left in March, citing “conflict over core priorities.” Sutskever followed in April. The safety team, now gutted, reports to a research VP who is evaluated on model performance, not alignment rigor.
Core: This is a systematic teardown of OpenAI’s safety governance, and I’ll expose it along four fault lines: independence, incentives, talent, and transparency.
First, independence. Safety auditing requires a structural separation from product development. That’s not my opinion; it’s a basic principle of risk management. In the ICO scandals of 2017, I watched projects fail because the auditor reported to the CFO who wanted the token sale to proceed. I wrote a technical report showing a reentrancy vulnerability in a Sydney-based token contract. The founders ignored it, patched it silently after I published my GitHub gist, and proceeded to raise $2.5 million. The same logic applies at OpenAI. When the safety team reports to the person responsible for shipping GPT-5, the incentive is to find reasons to approve, not to delay. The structural independence is gone.
Second, incentives. Research VPs are measured on benchmark improvements, deployment velocity, and user growth. Safety metrics are secondary. I recall analyzing the Ethereum gas wars in 2019 — liquidity providers optimized for yield, not for protocol health. They tolerated inefficient contract design because it maximized short-term fees. That same short-termism now governs OpenAI’s safety function. The research VP cannot afford to let safety team findings delay a model release, because his compensation is tied to shipping. Alignment research becomes an obstacle, not a feature.
Third, talent. The departure of Sutskever and Leike is not just the loss of two individuals. It’s the collapse of the “safety network” within the organization. I’ve spent years studying how institutional knowledge disappears when key dissenters leave. During the Terra Luna collapse, I modeled the death spiral three weeks before it happened. I published a 20-page whitepaper with the mathematical proof. It was ignored because the authors of the seigniorage model were still in charge. The same is happening here: the people who understood the deepest alignment risks are gone. The people who remain are incentivized to optimize for metrics that look good on a dashboard.
Fourth, transparency. The new structure is opaque. No public statement from OpenAI explains how safety decisions will be escalated to the board. No details on whether the new VP has the authority to halt a deployment. No independent audit was published. In crypto, we call this “washing the signal.” When a protocol quietly changes its multisig without a community vote, you know governance is broken. This is the same signal. The ledger remembers what the mempool forgets. The organizational chart may change, but the inherent conflict remains.
The contrarian angle: Bulls will argue that this restructuring improves efficiency. They’ll point out that many top AI labs have safety teams embedded within research. They’ll claim that Leike and Sutskever were slowing down progress with excessive caution. There is some truth to this. OpenAI’s competitors — Google DeepMind, Anthropic — also have safety functions integrated into research departments, and they ship models. But there is a difference: those labs maintain independent safety review boards or external advisory panels. Anthropic’s Constitution is not just a paper — it’s enforced by a separate governance layer. DeepMind has a Ethics & Society review board. OpenAI dismantled its parallel structure without creating a new one. That’s the unforgivable sin.
Another contrarian view: maybe the reorganization will let the safety team access more compute and more data, leading to better empirical alignment. But that’s a tool, not a structure. Better tools won’t fix corrupted incentives. I’ve seen this in the NFT floor price manipulation scandal of 2021. Projects with high floor prices were washing trades through clustered wallets. The data was there, but the incentives to investigate were missing because the “floor price” was the metric everyone wanted to pump. The illusion persists until the liquidity dries. OpenAI’s liquidity is its talent and its trust. Both are draining.
Takeaway: The lesson for every AI stakeholder is clear: code is not law, it is merely preference. And the preference of OpenAI’s current leadership is speed over safety. Truth is a derivative of transparent data, and right now the data says that governance independence is dead. If you are building on OpenAI’s APIs for a regulated industry, you are now exposed to a risk that cannot be mitigated by monitoring model outputs. The risk is architectural. The restructuring has removed the single point of failure, but replaced it with a system designed to fail silently.
I’ve been writing about crypto’s governance failures for years. The patterns are identical: first the narrative shifts, then the personnel changes, then the safety audit becomes a rubber stamp, and then — if we are unlucky — the bug that kills the model is discovered after deployment. OpenAI has traded long-term trust for short-term velocity. The ledger remembers. The market will price this risk eventually. The only question is whether the correction comes before or after the incident.