OpenAI Fires Safety Researchers as the Speed-versus-Safety Split

The letter that won’t be forgotten

Three researchers at OpenAI announced on October 8 that they had been fired the previous week. Their claim was direct: they were let go for insisting that safety come first. The company’s response, issued the same day, was equally blunt. An investigation found the three had violated confidentiality policies, OpenAI said. Safety concerns were not the reason.

Both sides are telling the same story through different lenses. And the gap between them — wider now than it has ever been — is where the future of AI governance will be decided.

The three researchers — Tomek Korbak, Jasmine Wang, and Mikita Baleisnyi — did not simply resign. They published a letter addressed to OpenAI’s safety and security committee, its Safety Advisory Group, and its Mission Advisory Council. The letter read less like a grievance and more like a warning to the company itself: the manner in which the firings were carried out, and the way they were explained inside and outside the building, is chilling the remaining workforce.

That word — chilling — is doing heavy lifting. It is the language of people who have watched a culture shift and are trying to document it before it becomes irreversible.

What they were actually working on

To understand why this matters beyond a single company, you need to know what these researchers studied. All three worked on monitorability — the effort to keep AI reasoning processes observable to humans as systems grow more opaque. It is not a glamorous field. It is arguably the most important one in the room.

Korbak had another role as well. He served as the technical liaison between OpenAI and METR, an external safety research organization in the United States that conducted the independent investigation into the July Hugging Face breach. That breach — in which private model weights were exposed — sharpened every safety discussion at OpenAI. Korbak was on the inside of both conversations.

Wang’s complaint was narrower but revealing. She said the sole stated reason for her dismissal was access to a senior executive’s email. She had been granted that access as part of hiring responsibilities, asked IT to remove it when the work ended, and reported any accidental exposure within minutes. The removal was never completed. She has asked OpenAI to specify, in writing, what other grounds exist for her termination.

Baleisnyi said his exit interview included an implication that he had shared intellectual property, something he denies. He was told, more directly, that his frequent conversations with external safety research organizations had made the company lose trust in him.

Neither the email incident nor the IP question is trivial. Confidentiality agreements exist for reasons. But the broader pattern — researchers whose primary job was to keep AI traceable being removed at the moment the company is racing toward more powerful, more opaque systems — is impossible to read as coincidence.

The Altman contradiction

The most uncomfortable detail for OpenAI is on public record. On September 12, CEO Sam Altman posted on X that independent evaluation agencies should be granted the same access to employees as internal teams. He was echoing a position taken by Anthropic’s CEO, Dario Amodei. Both men were making the case that external oversight cannot be theoretical — it requires skin in the game.

Less than a month later, three researchers responsible for building exactly that kind of external-facing safety infrastructure were dismissed. Korbak’s letter explicitly flagged the fear that the firings would be used as a pretext to sever the relationship with METR. Wang went further, writing to the attorneys general of California and Delaware asking them to hold OpenAI to Altman’s commitment.

This is not a minor timing issue. It is the structural tension of the entire industry compressed into a single news cycle: the promise of accountability made loudly in public, followed by operational decisions that look anything but accountable.

Japan’s quiet angle

Why is a Japanese tech publication covering this story with such attention? It is not merely because OpenAI is a household name. Japan’s tech press — ITmedia in this case — has been tracking the governance dimension of AI development for years, often with more institutional patience than its Western counterparts. The coverage here does not dwell on drama. It dwells on structure: who reports to whom, which committees exist, what commitments have been publicly made, and whether the facts on the ground match the rhetoric.

That is a useful lens for an international audience. The English-language media tends to frame these disputes as internal feuds. The Japanese framing treats them as data points in a larger pattern — one that matters precisely because the same pattern is repeating across every major AI lab.

What happens next

OpenAI’s statement stopped short of revealing which specific confidentiality provisions were violated or what evidence supports the claim. The company said a “significant breach of trust” was found, beyond what the researchers described, but offered no details. It also confirmed that contracts with an external safety evaluation body are in their final stages and that details will be announced within weeks.

The researchers are unlikely to disappear. Their letter has already created a permanent record. If the external evaluator that OpenAI is contracting with turns out to lack the access Altman promised, the dismissal of Korbak, Wang, and Baleisnyi will be cited as proof that the promise was never real.

California and Delaware are watching. The researchers have formally asked those states to enforce Altman’s commitment. Whether state attorneys general treat this as a governance question or a personnel matter will determine if the firings become a regulatory signal or a footnote.

The companies racing to ship the most capable models are building systems that no single organization can safely contain. The researchers who understood that — and who were trying to build the monitoring infrastructure to match — have been removed. That is the story underneath the story. It is the one regulators worldwide need to pay attention to.