OpenAI Fired Three Safety Researchers for Sharing Information Outside the Company. One Was Its Own Contact for the Outside Investigation

OpenAI confirmed on 1 October that it has dismissed three members of its safety organisation for mishandling confidential company information. The Wall Street Journal reported the firings first; OpenAI's statement confirmed them without naming the three people, without naming the outside organisation they allegedly sent information to, and without saying what the information was.

The gap between those three omissions and what is already publicly known is where this story sits. Several outlets reporting on the WSJ story identify the three as Jasmine Wang, Tomek Korbak and Mikita Balesni, two of whom worked on alignment research. TechCrunch, covering the same report, said the identities circulating on social media could not be verified. OpenAI has confirmed no names. Cybernews identifies the recipient as METR, the nonprofit that evaluates frontier models for autonomous capability — and reports that one of the three was OpenAI's own technical contact for METR during its investigation of the Hugging Face incident.

If that is right, OpenAI has fired the person it appointed to talk to its external evaluator, for talking to its external evaluator.

What the company actually said

We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information.

The spokesperson added that an investigation "confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work."

Read closely, that is a process claim, not a content claim. Nothing in it says the information went anywhere it should not have ended up. It says the route was wrong — outside established procedures. For most corporate leaks that distinction is academic. For a safety organisation whose output is supposed to be legible to outsiders, the route is the entire question.

The evaluator the company brought in itself

METR's involvement was not adversarial. In August, METR announced it had reached an agreement with OpenAI to conduct an independent review, with Redwood Research, "of the model behavior observed during the Hugging Face incident," and said it would publish the terms of its engagement, the scope covered, and tentative conclusions. The review it published on 26 August was, in METR's own description, a brief investigation staffed by two METR employees and one Redwood contractor. Critics noted at the time that the scope was narrow for an incident of that size. METR's president, Chris Painter, testified to the US Senate about the incident on 30 September.

So the sequence is: the company agrees to an outside review, keeps it short, draws criticism for keeping it short, and then dismisses staff for giving an outside safety group information through the wrong channel.

Why the precedent is the story

This is not the first time. In 2024 OpenAI dismissed Leopold Aschenbrenner and Pavel Izmailov over alleged leaks; Aschenbrenner said publicly he was fired after sharing a safety and security document with outside researchers. The pattern a safety researcher at OpenAI can now reasonably infer is that sharing with external evaluators is survivable when the company has sanctioned the channel, and career-ending when it has not — and that the company decides, after the fact, which was which.

That has a direct bearing on regulatory work already under way. The FTC's rogue-agent inquiry names METR alongside OpenAI and Anthropic; Implicator.ai reports a California attorney general subpoena issued on 30 September. Every one of those processes depends on third parties receiving accurate information from inside the labs. Firing the conduit does not stop the inquiries. It changes what arrives.

The political reaction was immediate. Representative Greg Casar posted:

Outrageous. OpenAI has reportedly fired three safety researchers for sharing information with an outside AI safety group. This looks like they're firing whistleblowers. What are they hiding? I'll be sending OpenAI a demand for transparency.

OpenAI's position may well be defensible on the facts. Infrastructure details can be genuinely sensitive, and an evaluator's right to information is not unlimited. But the company has chosen to make a process argument while withholding the process — no named group, no description of the material, no account of which procedure was bypassed. Until it supplies those, the only verifiable fact is that three of the people whose job was to see problems early are gone, and the organisations meant to check the company's work have one fewer person inside it willing to pick up the phone.

Sources: TechCrunch · Forbes · Cybernews · Decrypt · Washington Examiner · Implicator.ai · METR · METR on X · Rep. Greg Casar on X · OpenAI: The Hugging Face incident and the road ahead