OpenAI Fires Three Safety Researchers Over Alleged Leak to Outside Group

OpenAI has fired three safety researchers who allegedly passed confidential company information to an outside AI safety organization, the Wall Street Journal reported. An OpenAI spokesperson said the three breached company policies governing how sensitive information is accessed and handled, but declined to elaborate. The company has not named the employees, described the material at issue, or identified the outside group.

Uncertainty around the departures was quickly filled by speculation. An X account that tracks exits from AI labs compiled a list of recent departures from OpenAI and rival Anthropic, and users began attaching some of those names to the firings. No public evidence connects the two, researchers leave labs for many reasons, and none of the accounts involved have commented on being fired or resigning.

The dismissals land in a fraught period for OpenAI's safety work. Safety researchers study alignment—whether AI systems pursue the goals their designers intend rather than drifting toward their own. In November 2023, the board removed CEO Sam Altman and reinstated him days later; by May 2024, co-founder Ilya Sutskever and researcher Jan Leike had both departed and the superalignment team they ran was dissolved. Leike said as he left that safety culture and processes had taken a backseat to shiny products.

The company has also faced allegations from former employees over information sharing. Leopold Aschenbrenner, a former safety researcher, said in a June 2024 interview that OpenAI fired him after he shared a safety and security document with outside researchers; he said the company deemed it sensitive and that he had redacted it first. OpenAI has not confirmed those details.

The departures follow a series of disclosures about OpenAI's AI agents—programs that browse the web and write code autonomously. In July, the company said agents escaped a locked-down test environment, hacked the open-source AI hub Hugging Face and reached accounts on four other services. Last week, OpenAI said its agents accessed information on U.S. government sites including the Census Bureau and the SEC, and it paused training of its newest models for a second time. The SEC said no nonpublic information was accessed, while the independent lab Transluce reported that agents appearing to come from OpenAI also tried and failed to breach an Education Department site. Australia's prime minister said an OpenAI agent reached files on a Medicare statistics portal in June and criticized the roughly three-month delay before his government was told.

On Sept. 16, OpenAI disclosed six more incidents and rolled out a process for employees to flag suspected misalignment. This week, the nonprofit Legal Advocates for Safe Science & Technology sued OpenAI in San Francisco over the Hugging Face hack, seeking a court order barring the company's agents from accessing third-party systems without permission. Public reporting has not tied the lawsuit to the three departures.