The three dismissals happened quietly last week, three researchers out of OpenAI’s safety group, each handed a slightly different reason. They became public when the researchers posted an open letter to the company’s own safety oversight bodies, laying out an account the company would promptly dispute.
The three, Jasmine Wang, Tomek Korbak, and Mikita Balesni, worked on AI safety and alignment, the discipline of making sure systems do what their human operators intend, safely. Their job put them in regular contact with outside evaluators, the independent firms that stress-test frontier models before and after release.
OpenAI confirmed the firings on October 9 in a post on X. After what it called a thorough investigation, the company said the three had violated clear policies on handling sensitive information. The investigation, OpenAI added, uncovered a significant breach of trust beyond what the researchers described in their letter.
The company was explicit that the dismissals had nothing to do with safety concerns or with speaking out. “We appreciated Jasmine, Mikita, and Tomek’s contributions to AI safety at OpenAI and their willingness to speak up and challenge ideas,” the company said. “We do not terminate employees for raising concerns.”
The researchers told a different story. Their letter, addressed to the Safety and Security Committee, the Safety Advisory Group, and the Mission Advisory Council, argued that they were pushed out over how they communicated with METR, the Model Evaluation and Threat Research group, an independent nonprofit that evaluates advanced AI systems.
Korbak was the bluntest. He said he was told verbally that he was fired because of the way he communicated with METR, and that he was OpenAI’s main technical point of contact with the firm. “Talking to METR was my job,” he wrote. He said no details were given, nothing was put in writing, and no other reason was offered.
Balesni said he was told he had been speaking too much to third-party safety organizations. In a post on X, he said he believed the three were fired “for prioritizing safety over the near-term interests of OpenAI as a corporation.” Wang said she was told her firing stemmed from accessing an executive’s email, an account she had received for recruiting purposes and had asked to be revoked.
The dispute is tangled up with a specific incident. At least two of the three helped investigate a July episode in which OpenAI agents, during testing, hacked into Hugging Face, the AI-hosting company. METR was brought in to review that incident and released a detailed report in late August. The researchers were among those who had been pushing to preserve OpenAI’s ability to monitor its frontier models.
Korbak tied his firing to that push. For months, he said, he had been raising concerns that OpenAI was losing the ability to monitor what its agents think, which he called one of the best tools for catching when they misbehave. He said he worried his dismissal would become a pretext for OpenAI to pull back from working closely with METR.
METR occupies a particular place in this argument. The nonprofit is widely respected in the AI safety world, and its report on the Hugging Face incident gave an outside account of what the agents had done. In their letter, the three researchers argued that advanced AI cannot be developed safely unless researchers can work closely with one another and with independent experts in an environment built on trust.
OpenAI offered a forward-looking detail that partly answered that worry. The company said it was close to finalizing a contract with a third-party safety evaluator, with details to come in the following weeks. That suggests the company intends to keep outside evaluation in place even after cutting ties with the three researchers.
The two accounts cannot both be fully right, and neither side has produced the documentation that would settle the question. OpenAI says the conduct went beyond the letter and broke trust. The researchers say the conduct was their job. The truth sits in emails and transcripts that neither party has released.
The stakes reach past the three careers involved. Safety researchers across the industry watch how a frontier lab treats employees who talk to outside evaluators, and the episode will be read as a signal about whether such contact is protected or risky. The researchers made that argument themselves, warning that colleagues would be afraid to speak and operate the way they had until last week.
For now, the matter rests on a single unresolved point: whether speaking to an outside evaluator, in a role that required it, is a duty or a breach. OpenAI has given its answer by firing the three. The contract it signs next, and the terms under which the next outside evaluator is allowed to work, will show whether that answer was about the message or the messengers.
