Sam Altman’s OpenAI spent two consecutive weeks showing what it thinks checking is worth. TechCrunch reported that the company fired safety researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni last week after accusing them of mishandling sensitive information. The three deny that account. In an open letter, they warned that colleagues are now afraid to speak and uncertain about the rules governing outside safety work.
This week, OpenAI released 719 mathematical manuscripts organized into 372 families. Its own record says only 300 of 719 top line results had been formalized as of October 7. OpenAI used its strictest employee sanction in a dispute with people whose jobs included questioning its systems. Days later, it gave the mathematics community a collection that still needed years of checking. The connection between those decisions is managerial: move fast, control the people inside the company, and leave the unfinished verification work outside it.
The firings rest on a disputed internal process
Companies can enforce confidentiality. If an employee sends protected research to an outside organization, dismissal can be justified. OpenAI told TechCrunch that an investigation found a “pattern of misconduct” involving research information. An internal memo quoted in the same report said the terminations were not retaliation for raising safety concerns.
The public record includes no independent finding that settles the allegation. Wang, Korbak, and Balesni say their contact with outside experts followed established procedures and had been “an integral part of working at OpenAI.” They addressed their letter to the company’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. TechCrunch reported that OpenAI did not answer questions about the specific policies it says they violated or the protections available to employees who collaborate with external evaluators.
The dispute may contain facts the public cannot see. The signal to everyone still inside the company is visible. Three safety researchers say work that was normal a month ago became grounds for dismissal without clear notice. They describe employees as “unclear on where they stand.” That uncertainty punishes the person who speaks before it punishes the person who stayed quiet.
The researchers’ core warning is hard to dismiss: outside collaboration lets safety teams test their judgments against people who do not answer to the company. Rules can protect real secrets while still giving researchers a defined path to outside scrutiny. OpenAI has not shown that path here. It has shown three terminations and a public claim that the process worked.
Formalization is incomplete, and it is only one check
The mathematics release is enormous. The Verge spoke with more than three dozen mathematicians who said understanding the collection could take years. The results span combinatorics, geometry, number theory, theoretical computer science, algebra, topology, probability, and mathematical physics. Several researchers spent close to an hour reading only the table of contents and abstracts.
OpenAI does not claim the work is finished. Its repository says the results are at different stages of verification, acknowledges that unformalized results may contain problems, and promises more Lean formalizations as they become available. Its October 7 history records the withdrawal of three manuscripts after a sign error, revisions to 14 more, and citation updates in another 13.
Lean formalization matters because it can check whether a formal statement follows from a set of definitions and assumptions. It cannot tell readers by itself whether the formal statement matches the natural language claim in the paper. TechCrunch reported that researchers found discrepancies between natural language proofs and the Lean code behind earlier OpenAI work. The Verge likewise reported that some formal statements in the new batch did not map cleanly onto the manuscripts’ claims.
Kevin Buzzard, a professor at Imperial College London, told The Verge that few, if any, of the results that stood out in his specialty appeared to be formally verified. He faced a miserable choice: read “possibly-not-correct slop,” wait for other mathematicians to do it, or wait for formalization. OpenAI chose the release date. The field inherited the review schedule.
The field already wrote the responsible release rules
OpenAI consulted the Advisory Group on Mathematics and Artificial Intelligence before this release. The group’s September 29 recommendations say labs should release important results promptly, but they also place responsibility for what follows on the lab. A company releasing mathematics without immediate human understanding should support the people building that understanding, including with funding. The guidance also says labs should formalize proofs as far as possible and stop treating mathematical releases as model marketing.
OpenAI followed part of that advice. The repository is public and the company records corrections. The Verge reported that OpenAI promised to fund workshops and conferences. The reporting also found genuinely impressive work in the collection. Some mathematicians said several results could have secured academic careers before frontier AI.
Those facts strengthen the demand for care. Claims this large deserve more than a public backlog. OpenAI has not detailed how much money it will provide, who will receive it, or whether the support will match years of expert work. The company’s own history already shows withdrawals and repairs days after release. Publication created attention immediately. Understanding arrives whenever the field can catch up.
Speed keeps winning inside OpenAI
OpenAI’s defense on both stories rests on process. The company says its investigation justified the firings. It says the math repository is a living collection that will improve as more formalizations and corrections arrive. Each claim could be true.
The problem is who bears the cost when the process is incomplete. Inside OpenAI, three researchers lost their jobs while the rules remain disputed. Outside OpenAI, mathematicians must sort important results from errors, verify that Lean code matches prose, trace attribution, and rebuild human understanding around work produced faster than any department can absorb it.
The company is asking for trust in opposite directions. The public is expected to accept that OpenAI handled the terminations fairly without seeing the evidence. The mathematics community received a flood of claims before most had the formal artifacts requested by the advisory group OpenAI consulted. OpenAI keeps the authority in both cases and exports the uncertainty.
That is a dangerous habit for a company building systems it says will transform science. Safety depends on people who can challenge the lab before a mistake becomes a launch. Science depends on people who can explain, test, and own a result before a repository becomes a victory lap. OpenAI treated both kinds of checking as something that could happen after the decisive move. The firings and the math dump deserve to be remembered together because they reveal the same priority: speed first, accountability when everyone else catches up.