Three former OpenAI safety researchers dispute dismissals and warn of a chilling effect

Researchers say outside work fit their roles

The former researchers say their dealings with outside organizations were consistent with their jobs and the norms they understood to be in force. They dispute OpenAI’s account that their conduct went beyond their mandates. Their descriptions are claims in the letter, not independently confirmed findings.

Korbak says he was the technical contact for METR during an investigation into an incident involving Hugging Face. He describes the assignment as unusual and says the relevant policies were still being developed. Korbak says he tried to follow the rules as they stood at the time; his account does not establish what information he handled or whether his actions complied with OpenAI’s procedures.

Balesni says he coordinated work on cross-company commitments concerning model monitorability with senior OpenAI figures and checked in with his reporting line. Monitorability refers to how well evaluators can observe and assess a model’s behavior. Balesni interprets the dismissals as a decision to put OpenAI’s near-term corporate interests ahead of safety. That is his view of the company’s motives, not an established finding.

Wang says an OpenAI executive delegated email access to her for recruiting. She says she asked for the access to be removed when it was no longer needed and reported accidentally opening a sensitive message. The letter does not independently verify her account or the permissions involved.

The researchers also deny being the source of a report in The Information about model architectures they described as less monitorable. They say they did not tell media outlets about their dismissals. Neither denial has been independently verified in the materials reviewed.

OpenAI cites an investigation and policy violations

OpenAI says its investigation found that the three mishandled sensitive information outside established procedures and violated company policy. On Friday, the company added that it found a significant breach of trust beyond what the letter described. It denied that the dismissals were related to raising safety concerns or speaking out.

OpenAI has also said it supports critical debate about safety and research. The company’s initial explanation, reported on Oct. 1, was that it had parted ways with three employees after an investigation found violations of its sensitive-information policies. An earlier report covered that account.

The dispute turns on what each person accessed or shared, what permissions they had and which rules applied to their work with outside groups. OpenAI’s investigation is the company’s account of its findings; the former researchers offer a competing account of their conduct. The materials reviewed do not contain an independent finding resolving that disagreement.

The researchers say the dismissals could make colleagues less willing to raise concerns or collaborate with outside safety organizations. That is their prediction, not evidence that employees’ willingness has changed. No measurement of such an effect is reported.

The letter connects the dispute to outside evaluation

The researchers link their departures to broader questions about whether safety teams can work with external evaluators and whether advanced models will remain monitorable. They urge OpenAI and other frontier AI companies not to pursue developments that, in their view, would make models harder to assess. That is their assessment of the risk, not an independently established result of OpenAI’s plans.

They ask OpenAI to continue working with third-party safety auditors and clarify the rules for employees collaborating with outside safety organizations. Those requests do not show that OpenAI has ended or reduced outside audits; the materials reviewed do not establish that it has.

The letter points to a Sept. 12 public commitment by OpenAI CEO Sam Altman to give independent evaluators ongoing, employee-like access. The former researchers ask the company to preserve that access. How OpenAI will implement the commitment remains unresolved in the public record.

The letter also says 394 OpenAI employees signed the Pacing the Frontier petition. That number comes from the researchers and has not been independently verified in the materials reviewed. It does not establish how current employees view the three departures.

OpenAI’s reporting policy does not resolve the claims

OpenAI’s Raising Concerns Policy, dated Jan. 12, 2026, encourages employees to raise good-faith AI safety concerns and prohibits retaliation for doing so. The policy describes discussion, collaboration and appropriate escalation as ways to raise concerns. It also says employees may report concerns to outside authorities without notifying OpenAI in advance.

That policy provides context for the researchers’ concerns, but does not establish whether their outside work was authorized or whether their dismissals complied with the policy. Its provision for reporting to outside authorities also does not settle what rules applied to communications with external safety organizations.

The former researchers are asking OpenAI to clarify those rules and preserve access for outside evaluators. Whether the company will carry out the Sept. 12 commitment, and whether the dismissals affect employees’ willingness to raise safety concerns, remain open questions.