Fired OpenAI safety researchers give their first interview: they deny leaking and say Signal was part of the job

On October 23, the NPR podcast Machine Gods aired the first interview given by Mikita Balesni and Tomek Korbak since OpenAI fired them alongside Jasmine Wang in early October; Wang declined to appear. Balesni worked on alignment evaluations and reasoning monitoring at OpenAI after co-founding Apollo Research; Korbak researched chain-of-thought monitorability and served as OpenAI’s point of contact with METR during the Hugging Face incident investigation. Over roughly 53 minutes, the episode laid out the firing’s trigger for the first time: a week before their dismissal, the company asked them to hand over Signal messages on their personal phones, and found disappearing messages turned on.

Key points

• Balesni said on the show that a week before the dismissal, all three were pulled into a leak investigation over The Information’s report on less monitorable architectures. They denied being the source and, as asked, showed their personal-phone Signal messages with an outside researcher.
• Korbak called Signal use with third parties “standard operating procedure”; both had disappearing messages on. OpenAI told The Wall Street Journal the firings were for violating “clear policies for handling sensitive information” and were “not for raising safety concerns or speaking out.”
• Korbak’s role as METR’s point of contact is verifiable in the public letter PDF; by his account, METR planned a two-day visit that grew into six days across three trips.

The firing trigger

The timeline comes from two kinds of sources: the researchers’ own account on NPR, and The Wall Street Journal’s reconstruction — which sits behind a paywall (HTTP 403 when we tried), so it reaches this page only through The Verge and the Chinese outlet IT之家, attributed at each step. The starting point was The Information’s report on OpenAI’s Astra model: more capable, but exposing less of its reasoning. IT之家, citing the Journal, dates the piece to September 1; NPR gives no date, so the September 1 claim rests on that one attribution. Per the same account, OpenAI’s security team suspected the three had passed internal architecture details to an outside researcher, asked them to unlock their phones, found Signal chats with disappearing messages enabled — and fired all three the same day a week later.

On NPR, Balesni kept the recipient unnamed — “they specifically thought the person who leaked was an external safety researcher that we talked to a lot,” he said, adding that all three handed over the records as asked. The name comes from the Journal’s reporting: per The Verge and IT之家, the researcher is Ryan Greenblatt, formerly Redwood Research’s chief scientist, who has denied leaking anything to journalists.

The mechanics of the dismissal got detail too. An hour before the meeting where they were fired, Balesni recalled, a message arrived: “We have a meeting with HR.” When he asked to explain the context, he was told it was a final decision. Korbak says he still has no written explanation — a post by OpenAI’s research leaders on X is “the best explanation I was given.”

Signal sits at the center of the dispute. Most METR communication ran on OpenAI’s internal Slack, Korbak said, with some over email and Signal — and Signal was what worried the security team. He called it standard operating procedure (“I believe essentially everyone is doing that at OpenAI”), and Balesni noted his team had a Signal group chat with people from other companies and third-party organizations. Both had disappearing messages on. The security team, Korbak said, wanted to photograph the conversations; for the two researchers, auto-deletion is just good security hygiene. “When you cannot see evidence of guilt, you’d presume innocence. But somehow at OpenAI, when you are not able to prove your innocence, you are presumed guilty.”

Working with METR

Korbak’s role in the investigation is on the record: the open letter PDF describes him as “the technical point of contact for METR in the Hugging Face incident investigation.” On the show, Balesni went further — without Korbak, he said, the METR investigation probably would not have happened. Korbak described the engagement’s shape: METR was meant to be on site for two days, then invited back twice as the incident’s scale kept growing, for six days in total. Their mandate was a list of questions and a list of data sources — a scoping he called structurally flawed, because nobody knows in advance what an auditor should ask. His ideal is an auditor embedded like an employee, with unrestricted access. Management’s worry, he said, was legal: an evaluator could produce a report that state attorneys general would find damning.

Balesni’s outside contact ran on a different track: with Jasmine Wang, he pushed for cross-industry monitorability standards, writing internal memos that went to the board and senior executives. “I never shared any confidential IP with people outside of OpenAI,” he said. The two also described a chilling effect — colleagues now afraid to talk to them, to third parties, or to let their phones be searched. Rumors that OpenAI intends to cut off METR, they noted themselves, cannot be verified.

What others say

OpenAI’s responses so far are all secondhand. NPR says its request for comment went unanswered; host Kevin Roose read out the statement OpenAI’s research leaders posted on X — sensitive-information policy violations, and a “significant breach of trust” beyond what the letter describes — alongside the company’s denial that anyone was fired for raising safety concerns. To the Journal, OpenAI said the researchers violated “clear policies for handling sensitive information” and were “not for raising safety concerns or speaking out” (via The Verge). IT之家’s summary adds one layer: the company reportedly called Signal a non-approved channel, while stressing it was not the firing reason.

The three told the Journal the company never told them in writing what information was mishandled, and that they worry independent auditors will now think twice. Two discrepancies are worth logging: the September 1 date for The Information’s piece currently traces only to IT之家, and Greenblatt’s role differs — The Verge calls him Redwood Research’s former chief scientist, while IT之家 says he was commissioned by METR for the investigation. Those can coexist: in the assessment we covered on October 22, METR and Redwood each sent one researcher.

The interview’s increment is easy to state. The October 8 open letter denied the misconduct allegations point by point and never mentions Signal — the disappearing-messages detail is new. The three-week METR and Redwood assessment went public on October 22 with OpenAI holding redaction rights; before all this, OpenAI confirmed the firings on October 1 without naming anyone.

Our take

The month tells the story better than any single event: OpenAI hosted METR and Redwood for a monitoring assessment while firing the three safety researchers most deeply involved in external evaluation. Balesni argued on the show that voluntary engagement is unsustainable — auditors in other industries are mandated by law, can look at everything and talk to anyone, while AI, with higher stakes, relies on a few insiders pushing from within. Korbak added the structural point: frontier labs lack the incentives for checks and balances, because leadership’s fiduciary duty is to revenue. Both men are arguing their own case, and the underlying question — what counts as compliant communication with external evaluators — does not go away because they left.

Disappearing messages deserve their own note. For security practitioners, end-to-end encryption plus auto-deletion is routine; for an employer, it is evidence that ceases to exist. That conflict used to stay inside industry group chats. Here it ended three jobs, at a company that had publicly promised embedded auditors. The honest caveats: no written firing rationale has surfaced, the “significant breach of trust” is unspecified, the METR rumors are unverifiable, and OpenAI had not responded to NPR by airtime. What can be checked independently is the letter PDF, the NPR transcript, and the two on-record lines OpenAI gave the Journal.