OpenAI Says Three Researchers Violated 'Clear Policies.' It Has Not Said Which.

Two primary documents exist in the dispute over three fired OpenAI safety researchers, and they were published about a day apart. On 8 October Tomek Korbak, Jasmine Wang and Mikita Balesni released a four-page open letter addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. On 9 October the company’s newsroom account on X posted “A note from our research leaders”. Most of what has been written since is coverage of those two texts.

I read both, and I read the news reports that quote them. My conclusion is narrower than the headlines: as of 14 October 2026, I found no document from either side that names the policy at issue, confirms the information involved, or states the date a rule took effect. The two accounts are not reconciled by any public fact. They simply do not meet.

What is on the record, and who said it

The Wall Street Journal reported the departures on 1 October. Its article is paywalled, so I am relying on TechCrunch’s summary of it, which quotes an OpenAI spokesperson: “We have parted ways with three individuals” for violating policies on accessing and handling sensitive company information, and the investigation “confirmed that these individuals mishandled sensitive information outside established company procedures.” TechCrunch noted that the WSJ report did not name the researchers, the outside organization, or the information.

The firings happened “last week,” in the researchers’ words on 8 October, so the exact day is not published. I have drawn that gap as a dashed band below. The band runs from about 28 September (the start of the calendar week before the 8 October posts) to the 1 October report; those edges are my inference, not a published date.

Dated on-record events around the OpenAI firings, 12 September to 10 October 2026 A timeline on a horizontal axis from 12 September to 10 October 2026. 12 September: the company's public commitment on embedded evaluators, dated by the researchers' letter. 29 September: the New York Times reports warnings brushed aside, as summarized by TechCrunch. A dashed band, inferred, from about 28 September to 1 October marks the firings, whose exact day has not been published. 1 October: the Wall Street Journal reports the departures, via TechCrunch. 8 October: the researchers' open letter and three posts. 9 October: OpenAI's note from its research leaders. DATED ON-RECORD EVENTS · 12 SEP TO 10 OCT 2026 12 Sep 19 Sep 26 Sep 3 Oct 10 Oct Firings: day not published; "last week" per the 8 Oct posts, reported by 1 Oct firings: day not published, band inferred 12 Sep: the company's public commitment on embedded evaluators, date per the researchers' letter 29 Sep: New York Times report, as summarized by TechCrunch 1 Oct: WSJ report, via TechCrunch 8 Oct: researchers' open letter and posts 9 Oct: OpenAI Newsroom note 12 Sep Company pledge on embedded evaluators (date per the letter) 29 Sep NYT: warnings brushed aside (via TechCrunch) 1 Oct WSJ report, via TechCrunch 8 Oct Open letter and three posts 9 Oct OpenAI note: "violated clear policies"
Five dated texts, two of them primary: the researchers' letter (8 Oct) and OpenAI's note (9 Oct). The firings themselves sit in a gap that neither side has dated.

What OpenAI’s note says, and what it does not

The note is a company claim, posted by the OpenAI Newsroom account under the heading “A note from our research leaders,” with no individual named as author. It says the three were let go “after a thorough investigation found they violated clear policies on handling sensitive information.” It adds that the investigation “uncovered a significant breach of trust beyond what’s outlined in the letter they published,” and that OpenAI stands by the decision. It also says: “We have not and do not terminate any of our employees for raising concerns.” And it says the company is “actively finalizing contracts with third-party safety assessors,” to be announced “in the coming weeks.”

Read as evidence, that is three assertions with no checkable content: a policy that is “clear,” a breach that is “significant,” and a part of that breach that is, by the note’s own wording, outside the letter. The note also says OpenAI generally keeps individual employment matters private. That is a normal position. It also means the company’s best evidence is, by choice, not in the public record.

Fortune reported on 9 October that OpenAI said it found a “pattern of misconduct” and did not link the firings to the researchers’ communications with METR or to the Hugging Face incident. METR, the outside evaluator involved, did not respond to Fortune’s request for comment, and I found no METR statement of its own as of 14 October.

What the researchers say, also unverified

The letter is the researchers’ claim, and they are parties to the dispute. It makes four specific statements. They were not the source of the leak to The Information about “supposed new, less monitorable architectures,” and “we do not know who it was.” They do not believe they “engaged with external parties outside the mandates of our jobs,” because the Hugging Face investigation was “without precedent and internal policies were being developed in real time.” Balesni, the letter says, worked on cross-company commitments “in coordination and discussion with board members and the C-suite.” And Wang’s access to an executive’s email “was delegated for recruiting purposes, with permission”; she asked for it to be removed, IT did not remove it, and when she “accidentally clicked on a sensitive email” she reported it “within minutes.”

Their posts add process claims. Korbak wrote that he was told “verbally” he was fired because of the way he communicated with METR, with “no details on what I said or did or when,” and that “nothing was put in writing.” Wang wrote that she “was given one reason: that I accessed an executive’s email.” Fortune describes Wang as a program manager on OpenAI’s safety team; the letter calls the group “safety and alignment employees.” Korbak and Balesni also say they believe the firings were about safety concerns, which OpenAI’s note disputes. Those are two different stated reasons for two different people, which is consistent with, but does not show, the possibility that the company’s single statement covers three separate sets of facts. The documents do not say whether it does.

The researchers’ version also has holes. The letter does not say what Korbak told METR. It does not say which procedures existed at the time, only that they were still forming. A reader cannot check “within minutes” either; it is an unsupported timing claim until the log exists.

Five questions put to the two primary documents A two-column comparison of OpenAI's 9 October note and the researchers' 8 October letter and posts. Information involved: OpenAI says only "sensitive information"; Wang cites an executive's email and Korbak cites his METR communications. Policy: OpenAI says "clear policies" without naming one; the researchers say policies were being developed in real time during the Hugging Face investigation. Reason in writing: not stated in the note; Korbak says it was verbal only. Leak to The Information: not mentioned in the note; the researchers deny it and say they do not know who it was. Anything beyond the letter: OpenAI cites a significant breach of trust beyond what the letter outlines, unspecified; the researchers say a board-level memo rumor was never raised with them. FIVE QUESTIONS · OPENAI NOTE 9 OCT VS RESEARCHERS' LETTER AND POSTS 8 OCT Question OpenAI's note Researchers' letter and posts What was the information? "sensitive information" not described Wang: an executive's email; Korbak: his METR communications Which policy? "clear policies" none quoted or named Policies "were being developed in real time" Reason given in writing? not stated in the note Korbak: told "verbally," "nothing was put in writing" Leak to The Information? not mentioned in the note Denied: "we do not know who it was" Anything beyond the letter? "a significant breach of trust beyond what's outlined in the letter" content unspecified Board-level memo rumor "never raised with us"
Each cell is a quotation or a plain "not stated." No row I could find contains a fact that both columns assert and that one of them can contradict.

Why both can be true

OpenAI’s spokesperson said the three “mishandled sensitive information outside established company procedures.” The letter says internal policies were “being developed in real time” during an unprecedented investigation. Those sentences conflict only if “established” means “in force and communicated to these employees at the time.” Whether that is true is a question about a document and a date, and neither has been published. If a rule was written down on a particular day and the conduct came after, the company’s statement is straightforward. If the rule was written after, or never reached the people it was applied to, the letter’s “norms of the time” argument has weight.

Wang’s case is different in kind. An executive’s inbox that was combined with her own is a question for IT logs and delegation records. Again, nothing public settles it, and OpenAI’s note does not mention it. The letter says that the allegation of a “shared board-level memo” was never put to them. That leaves a symmetric situation: the company says there is more than the letter outlines, the researchers say they have not been told what it is, and an outside reader has only the sentences.

I would add one thing about the people. The letter says that before joining OpenAI, the three started a cross-industry position paper, “Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety”. The arXiv listing, posted in July 2025, has 41 authors, and the first two named are Korbak and Balesni. The letter separately quotes an OpenAI colleague’s public statement that chain-of-thought monitorability is “fragile and unfortunately trending in a negative direction.” OpenAI’s note agrees with the letter that preserving monitorability “requires an industry-wide commitment.” So the substantive technical question is one point where the two documents agree. The dispute is about process and trust, which is harder to measure than a benchmark.

What would settle it

I do not have a view on whose account is right, and the record does not support one. I can say what would change that:

  1. The text of the policy, with its effective date and how it was communicated to staff.
  2. The written termination reasons for each person, if they exist. Korbak says his were verbal only.
  3. What “a pattern of misconduct,” the phrase OpenAI’s spokesperson used with TechCrunch, refers to, and whether it is the same facts for all three.
  4. A statement from METR on what was said and under what arrangement.
  5. A reply from the three bodies the letter was addressed to. TechCrunch reported on 8 October that OpenAI had not formally responded to the letter; I found no public response from any of those committees as of 14 October.

The company has a fair point that it need not litigate personnel matters in public. It also has an unusual problem for a lab that published its own incident reports in detail last month, as I wrote in an earlier piece: the contrast between a report that names its failed assumptions and a note that names none is hard to miss. On the record today, the policy is unnamed and the reason is unwritten, or so one side says. I would wait for the document before taking either side’s word for what it says.

References

  1. Korbak, T., Wang, J., and Balesni, M. (2026). OpenAI cannot make AI safe on its own: A note on third-party collaborations, open debate, and clear operating procedures. Open letter, PDF, published 8 October 2026.
  2. OpenAI Newsroom (2026). A note from our research leaders. Post on X, 9 October 2026.
  3. Korbak, T. (2026). Post on X, 8 October 2026.
  4. Wang, J. (2026). Post on X, 8 October 2026.
  5. Mehta, A. (2026). OpenAI cuts ties with 3 safety researchers, WSJ reports. TechCrunch, 1 October 2026.
  6. Bellan, R. (2026). Fired OpenAI safety researchers dispute misconduct claims, warn of chilling effect. TechCrunch, 8 October 2026.
  7. Forlini, E. (2026). Controversy swirls at OpenAI over ‘abrupt’ firing of safety team members involved in the Hugging Face hack investigation. Fortune, 9 October 2026.
  8. Korbak, T., Balesni, M., Barnes, E., et al. (2025). Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety. arXiv:2507.11473, v1 15 July 2025.