Fired OpenAI Safety Researchers Deny Mishandling Confidential Information

Three fired OpenAI safety researchers say they did not mishandle sensitive information outside their normal job duties.

Jasmine Wang, Tomek Korbak and Mikita Balesni made that denial in an open letter published Oct. 8. The letter was addressed to OpenAI's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council, according to TechCrunch.

OpenAI says the three were dismissed for allegedly sharing confidential company information with an outside AI safety organization, and for accessing and handling sensitive information in violation of company policy. That account was first reported by the Wall Street Journal and is still the company's stated reason for the firings.

OpenAI shared with TechCrunch an internal memo from a research leader saying the dismissals were not about raising safety concerns or speaking out. The memo said OpenAI "does not terminate employees for raising concerns."

The researchers reject that description. In the letter, they said they did not work with outside groups beyond what their roles allowed. They also denied any role in a leak to The Information about newer OpenAI models using architectures that are harder to monitor, which makes chain-of-thought reasoning harder to track.

Chain-of-thought reasoning means the intermediate steps a model produces while working toward an answer. Safety teams read those steps as one check during evaluation, red-teaming, which is structured testing to find failures, and decisions about whether a model is ready to release. A design change that hides those steps would matter to staff who assess models before deployment.

The letter focuses less on model design and more on process. The authors wrote that firings like theirs, "executed and communicated abruptly, are chilling the open culture OpenAI has prized in the past." They said the way the firings were communicated has left former colleagues afraid to speak up and to work in ways that used to be normal at the company.

The firings were first reported on Oct. 1, when the Wall Street Journal reported that OpenAI had parted ways with three researchers over alleged misconduct including sharing confidential information with a third-party AI safety organization. That initial report was followed by coverage on Sept. 30 and Oct. 1 describing an internal investigation into mishandling of sensitive information.

Days later, a former OpenAI safety employee resigned and criticized the company's safety approach, saying the "time for trial and error is over," according to Reuters reporting on Oct. 3.

In my view, the argument is now less about access logs and more about where normal safety work ends and a policy violation begins. Frontier labs, the small group building the most advanced models, often have staff split time between internal testing, coordinating outside audits and talking with external groups. That only works if researchers know in advance which contacts are approved, which types of data can be shared, and how to get approval in unclear cases. When those rules are debated only after people are fired, trust breaks down quickly.

Worth flagging here is the follow-on effect described in the letter. Safety work depends on raising possible problems early and informally. If researchers respond by writing less down and collaborating less to avoid policy risk, reviews slow down and problems stay hidden longer. The longer trend still points toward better testing tools and clearer rules for outside sharing, which would protect confidential information without shutting down the internal discussion that catches problems before release.