OpenAI Fired Three Safety Researchers—and Neither Side Proved Its Full Case

On October 9, 2026, OpenAI confirmed it had fired three safety researchers—Jasmine Wang, Tomek Korbak and Mikita Balesni—and called the matter a “breach of trust.” The dismissals had taken place the previous week; the former employees publicly challenged them on October 8.

The dispute concerns whether the dismissals followed violations of information-handling rules or reflected hostility to safety work involving outside organizations. The employment decisions are confirmed. The evidence needed to establish either an individual policy breach or a retaliatory motive has not been made public.

The dismissals, the challenge and the reply

The sequence separates an agreed employment action from competing explanations. It also shows why the public statements cannot, on their own, settle what happened inside the investigation.

• Dismissals — confirmed action: Wang, Korbak and Balesni lost their jobs before either side published its detailed position. The precise grounds for each individual decision remain outside the public record.
• Researchers’ challenge — disputed claim: The former employees contested the handling of their dismissals and warned that uncertainty over permitted safety work could make colleagues afraid to speak openly. Their account describes a possible effect on staff, rather than a measured change in employee behavior.
• Company response — disputed claim: In its October 9 response, OpenAI said an investigation found violations of clear policies on handling sensitive information and a “significant breach of trust beyond what's outlined in the letter they published,” while denying that safety concerns or speaking out caused the firings.

Those positions leave a consequential gap. The company has identified a category of alleged misconduct but has not publicly specified the conduct behind each dismissal or the rules applied to it. The former employees have described why they believe their work was authorized, but their public account cannot establish the company’s motive. Neither side’s strongest conclusion follows simply from the fact that the researchers were fired.

Three accounts of safety work and information access

In their four-page letter to OpenAI’s safety oversight bodies, the researchers described the firings as “chilling the open culture OpenAI has prized in the past,” denied being the source of a leak about less monitorable model designs, and set out distinct accounts: Korbak was the technical contact for METR during an investigation involving Hugging Face; Balesni worked with senior leaders on cross-company efforts to preserve the ability to monitor model reasoning; and Wang had been granted access to an executive’s email for recruiting, sought to have that access removed and reported opening a sensitive message by mistake. These are the researchers’ accounts, not independent findings about their conduct.

Korbak’s case puts the boundary around external evaluation at issue. METR was involved in examining the Hugging Face incident, and communicating with an outside evaluator can be part of a safety researcher’s assignment. The unresolved question is whether the particular communications at issue stayed within the authority and information-sharing rules that applied at the time.

Balesni’s work presents a related boundary for collaboration across companies. Preserving the ability to inspect how advanced models reason may require researchers to compare concerns and possible commitments with people outside their employer. Such work still depends on decisions about what internal material may be shared, who can approve it and whether removing sensitive details is sufficient. The public accounts do not disclose the materials at issue or an approval record that would resolve those questions.

Wang’s account concerns a different route to sensitive information: access to an executive’s inbox rather than contact with an external evaluator. Her description distinguishes permission to access the inbox for recruiting from opening a particular message after that access was no longer needed. Whether the access request was acted on, how the message appeared in her mail app and what the investigation found about the incident remain undisclosed. Treating all three cases as the same alleged act would erase a material difference in the conduct being disputed.

What the investigation has made public

The company’s public explanation establishes that it investigated and chose to dismiss the researchers. It does not provide the policies, findings or individual reasoning needed to assess the allegation of a serious breach. A statement that the investigation found more than the former employees disclosed leaves the additional conduct unidentified for outside readers.

That limitation cuts both ways. Missing investigative details prevent a confident judgment that the dismissals were justified by clear, consistently applied rules. They also prevent a confident judgment that the stated reason was a pretext for punishing safety advocacy. The question is not whether either account is plausible; it is whether publicly available evidence distinguishes one explanation from the other.

The researchers’ warning about a chilling effect has a similarly defined status. Abrupt dismissals and uncertain rules could make other staff more cautious about raising concerns or contacting outside specialists. No public measure of staff behavior establishes how widely that has happened at OpenAI. The warning matters because it identifies a possible consequence for safety work, while the size and existence of that effect remain unverified.

The boundary for outside assessors

External safety assessment depends on researchers knowing what they may discuss and how to obtain permission to share sensitive material. An evaluator may need candid technical detail to examine a risk; the company may need to restrict information that could expose research or systems. Without a clear route between those needs, a researcher can face uncertainty even when external review is an accepted part of the work.

That governance question is sharper here because the disputed activities span outside evaluation, cross-company coordination and internal email access. A single broad reference to sensitive information does not explain which boundary was crossed in each case. Clear arrangements for independent assessors would show how future collaborations can proceed, but they would not resolve the individual dismissals. That would require evidence about the alleged breaches and the decisions made in the investigation.

Also read:

• Character.AI Chats Feel Private—Its Policy Still Treats Them as Platform Data
• The Mirage Effect: Stanford Just Proved That “Computer Vision” Is Often Just Confident Bullshit
• Volkswagen Enters the E-Bike Market with a Safety-Focused “Smart” Model — But at Car-Like Prices