OpenAI’s fired safety researchers say colleagues are afraid to speak

What the researchers dispute

The researchers—Wang, Korbak and Balesni—said they did not give The Information details about the architectures of OpenAI’s newer models, including claims that their reasoning chains are harder to monitor. They also said they had not worked with outside parties beyond the scope of their jobs.

Their letter argues that safety researchers need close contact with external experts to identify and address risks. Clear internal procedures matter, they said, but so does being able to raise concerns without fear. The researchers described their dismissals as part of a broader cultural shift: conduct they say was considered ordinary a month ago could now lead to termination.

The dispute also concerns two episodes the researchers say were connected to legitimate work:

OpenAI’s account leaves key details open

OpenAI did not formally respond to the open letter. It provided TechCrunch with an internal memo attributed to a research team leader. The memo recognized the three researchers’ contributions to AI safety and denied that they were fired in retaliation for criticism. It said their dismissals were unrelated to safety concerns or public statements, and that OpenAI encourages employees to raise such issues.

But the company did not answer TechCrunch’s questions about which rules the researchers allegedly broke, the circumstances of their dismissals or how employees who raise safety concerns and work with external evaluators are protected.

Wang gave a separate account on X. She said OpenAI told her she was fired for accessing an executive’s email. Wang said the access had been granted for recruiting work, that she had asked IT to remove it when it was no longer needed, and that the request had not been fulfilled. She said she could not revoke the access herself, and that the executive’s email appeared in her phone’s mail app in a way she could not distinguish from her own. After accidentally opening a confidential message, she said, she notified the executive within minutes and again asked IT to remove the access. She said she had hidden nothing.

The researchers’ call is for OpenAI to honor its public commitments: bring independent safety auditors into the company’s work, preserve the ability to monitor advanced models and sustain dialogue between its safety staff and the outside community. The internal memo, according to the letter, agreed with those recommendations.

The unanswered question is procedural

My read is that the central issue is not whether OpenAI says it supports safety criticism; the memo says it does. It is whether employees can know, in advance, where the line is when safety work involves sensitive information and outside experts. The company has not publicly described the rules at the center of these dismissals, while the researchers say those rules were unclear even as they acted.

That gap matters more when, as the researchers describe, an investigation is unfolding and the relevant procedures are still being worked out. If staff cannot tell which external contacts or access practices are allowed, the cost may be broader than a dispute over three dismissals: the people closest to emerging risks may decide that speaking up is not worth the uncertainty.