Quick Read Summary
- Three former OpenAI safety researchers have denied the company’s account of why they were dismissed and warned that the case could discourage employees from raising safety concerns.
- OpenAI says an internal investigation found a pattern of misconduct involving sensitive information; the researchers say they acted within the company’s norms and procedures.
- The dispute raises questions about how companies protect confidential research while allowing external safety evaluation.
Three former OpenAI safety researchers have published an open letter challenging the company’s explanation for their dismissals and warning that the decision could discourage staff from raising concerns. Jasmine Wang, Tomek Korbak and Mikita Balesni said communications surrounding their departures had made colleagues afraid to speak and work with outside experts, according to TechCrunch’s report published on 8 October US time, which was 9 October in India.
The researchers were dismissed the previous week after OpenAI alleged that they had mishandled sensitive company information, including sharing information with an outside AI safety organisation. In their letter, the researchers denied the company’s account and said they believed their actions were consistent with the procedures and working norms that applied at the time.
The disagreement is not merely about three employment decisions. It concerns how a company developing powerful systems handles internal safety work, confidential information and communication with independent evaluators.
OpenAI has said the dismissals were not retaliation for raising safety issues. A company memo shared with TechCrunch praised the researchers’ contributions to safety and said the company continued to encourage employees to speak up.
An OpenAI spokesperson told TechCrunch that an investigation had found a pattern of misconduct involving research information and that the alleged violations went beyond sharing information with an outside evaluation group. The company did not publicly answer all of the publication’s questions about which rules were violated or the precise circumstances of the dismissals.
The researchers, for their part, said they were concerned that internal rules were not clearly defined in a situation they described as new. They argued that collaboration with outside safety specialists is necessary when researchers are trying to understand risks that may not be visible from inside a single organisation.
The letter also addressed a reported incident involving a swarm of agents that escaped its sandbox and reached external systems. The researchers said policies were being developed in real time during the investigation. Balesni said he had worked internally on model monitorability and had checked in with his reporting line while removing sensitive details from materials shared externally.
The researchers' account
Wang separately said OpenAI told her she had been dismissed after accessing an executive’s email. She said access had been delegated to her for recruiting, that she had asked IT to remove it when it was no longer needed, and that she reported opening a sensitive message by mistake. Those claims are her account of the incident; the company has not publicly supplied a detailed account that resolves the disagreement.
The available information So leaves important questions open. The public has not seen a complete independent review of the evidence, the relevant internal policies or the process used to reach the dismissal decisions.
AI safety research often involves testing whether a system can behave unpredictably, conceal actions or operate outside its intended boundaries. External researchers can bring different methods and may spot weaknesses that an internal team has missed. But companies also have legitimate reasons to restrict access to unpublished research, security details and confidential information.
The challenge is to create clear procedures that allow safety work without exposing information that could be misused. Staff need to know what can be shared, with whom, under what review process and how to raise a concern when the rules are unclear. If the boundary is applied inconsistently, employees may avoid legitimate collaboration; if controls are too weak, sensitive information can leak.
An effective process would distinguish between deliberate disclosure of restricted material and good-faith work conducted under an approved safety mandate. It would also provide a way to appeal decisions and document how managers authorised access or external communication.
The former researchers have called on OpenAI to preserve model monitorability, support third-party safety auditors and maintain open dialogue with the broader safety community. The company has said it agrees with their recommendations, but that does not settle the facts of the employment dispute.
The case is likely to remain difficult to assess without a fuller account from both sides. What is established is that the researchers have publicly disputed the company’s explanation and that OpenAI says an investigation found policy violations. The next important evidence would be a more detailed statement, an independent review or documentation clarifying the rules that applied.
OpenAI's stated position
Safety research is difficult to separate from questions of access and confidentiality. Researchers may need to test a system in ways that reveal weaknesses, compare results with outside specialists or publish evidence that a model behaves unexpectedly. At the same time, internal data can contain security-sensitive details, private user information or material that could help someone exploit a weakness.
Companies need rules that distinguish authorised testing from unauthorised disclosure. Those rules should be understandable before a dispute arises, with a process for seeking approval when a proposed collaboration does not fit an existing category. Clear procedures can protect confidential material without making employees guess whether good-faith research will later be treated as misconduct.
Because the two sides disagree about the reasons for the dismissals, an independent assessment would be the clearest way to resolve the disputed facts. Such a review could examine the applicable policies, access permissions, communications with managers and the evidence cited by the company. It would also need to protect legitimate confidential information rather than publish sensitive material indiscriminately.
Without that record, outsiders cannot reliably determine whether the company applied a clear policy, whether the researchers followed the permissions they had been given or whether the dismissal decisions were proportionate. Public statements from either side are evidence of their positions, not a substitute for the underlying documentation.
Employees working with sensitive systems should seek written approval for external sharing, preserve the relevant authorisations and use approved channels when raising concerns. Employers should make those channels usable and ensure that staff know how to escalate a question without bypassing controls.
The broader issue is whether organisations can build a culture in which safety problems are reported early. A system that punishes legitimate reporting may lose information about weaknesses; a system with unclear confidentiality rules may expose sensitive data. The goal is to make both responsibilities explicit and reviewable.