OpenAI's Safety Exodus: Three More Researchers Gone as the Company Races Toward AGI

NewsOnScale Staff

October 2, 2026

When OpenAI was founded in 2015, its founding charter placed safety at the center of its mission — not as a department, but as a reason for the organization's existence. A decade later, the company has become one of the most commercially aggressive technology firms on the planet, and the people responsible for holding it to that original promise keep leaving.

The Wall Street Journal reported this week that OpenAI has parted ways with three safety researchers, adding to a growing list of departures that includes some of the most prominent voices in the field. The company has not publicly named the individuals or explained the circumstances of their exits.

## A Pattern, Not an Incident

This is not an isolated event. Over the past eighteen months, OpenAI has seen the departure of Ilya Sutskever, its co-founder and former chief scientist who led an internal safety-focused faction; Jan Leike, who ran the superalignment team and resigned with a pointed public statement about the company's priorities; and a string of researchers who worked on alignment, interpretability, and policy. Leike wrote on departure that safety culture had been "slowly eroding" and that capabilities work consistently won out over safety work in internal resource disputes.

Each departure was explained away individually. Collectively, they describe an organization whose safety infrastructure is being outpaced — and possibly outcompeted — by its commercial ambitions.

## What Safety Researchers Actually Do

It's worth being precise about what is being lost. Safety researchers at frontier AI labs are not simply ethicists writing memos. They work on technical problems: how to make models less likely to deceive users, how to detect dangerous capabilities before deployment, how to ensure that increasingly autonomous systems behave in ways their designers intend. This work is slow, often unglamorous, and produces results that are hard to ship in a quarterly earnings call.

The alternative — moving fast, releasing capable models, and patching problems after they emerge in the wild — is faster and more lucrative. It is also the approach that safety researchers exist specifically to resist.

When those researchers leave, the resistance leaves with them. What remains is a safety function that may be adequate for regulatory optics but increasingly thin as a technical check on the systems being built.

## The Accountability Gap

OpenAI occupies a structurally unusual position. It began as a nonprofit, converted to a "capped-profit" model, and is now pursuing a full for-profit restructuring. It has accepted billions from Microsoft and other investors. It is in active negotiations over valuation, product roadmaps, and market positioning that would make any serious safety friction extremely costly.

There is no independent board with meaningful oversight authority. There is no regulator in the United States with clear jurisdiction over frontier AI development. The company's own safety commitments are self-reported and self-evaluated. When the people doing that evaluation leave — or are shown the door — the accountability architecture becomes largely ceremonial.

## What OpenAI Says, and What It Doesn't

OpenAI has consistently maintained that safety remains a core priority, pointing to its ongoing investment in safety teams and its publication of safety research. Those claims deserve to be weighed against what is observable: the rate of departure, the seniority of those departing, and the public statements some have made on their way out.

Three researchers in another round of cuts may sound like a small number. But in a field where the total population of people capable of doing this work at the frontier is measured in the hundreds globally, every departure is significant. And in an organization that is simultaneously building increasingly autonomous AI agents, pursuing military and intelligence contracts, and racing competitors to capabilities benchmarks, the margin for safety infrastructure to thin any further is narrowing.

The question that OpenAI has not answered — and that no one with authority to demand an answer currently exists to ask — is a simple one: who, specifically, is responsible for ensuring these systems are safe, and what happens when they decide they're done?