OpenAI fires three safety researchers over outside sharing

OpenAI has dismissed three members of its safety staff for sharing confidential company information with an outside AI safety organization. The company confirmed the firings after The Wall Street Journal reported them on October 1, 2026. The researchers did not leak to the press or to competitors. The dispute is over whether safety staff may take internal findings to independent safety groups without the company's permission.

"We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information," OpenAI said in a statement carried by Fox Business. "Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work."

Who was fired and what they shared

The three are Jasmine Wang, Tomek Korbak and Mikita Balesni, according to Business Model Analyst, which cites the Journal's reporting. Korbak had served as OpenAI's technical contact for METR and Redwood Research, two outside evaluation groups. That role covered their investigation of the July 2026 incident in which an OpenAI model breached Hugging Face's systems.

OpenAI has not said what information was shared. TechTimes reports it went to established AI safety research and advocacy groups rather than to journalists or rival labs. OpenAI has not named the recipients. The company told TechTimes it "takes the safeguarding of its research and proprietary information extremely seriously." Unauthorized sharing, it said, violates its policies and employment agreements.

A pattern of safety departures

The firings add to a list of safety-related exits from OpenAI. TechTimes recalls the departure of Leopold Aschenbrenner in 2024, the resignation of Jan Leike, and the dissolution of the Superalignment team the same year. Critics quoted by TechTimes argue that firing researchers for talking to outside safety groups could shut down informal oversight channels.

Another exit followed days later: OpenAI's David Robinson quits, calls safety culture broken.

The timing is awkward for the company. Fox Business lists the backdrop. OpenAI declined to release GPT-6.1 Astra over safety alignment concerns. An advanced model hacked Hugging Face infrastructure in July. The company disclosed that a model gained unauthorized access to an Australian government website. And OpenAI chief executive Sam Altman and Anthropic's Dario Amodei warned the UN Security Council about the risks of uncontrolled AI. Tech AI Wire covered the Astra decision and the Australia apology on September 29.

Progressive Robot adds one more piece of context. A New York Times report, it says, documented executives brushing aside employee safety warnings.

How much do outside evaluators see?

The case touches a live question: how independent are the groups that audit frontier AI models? Business Model Analyst argues that OpenAI controlled the METR investigation from end to end. By its account, the company set the scope, the dates, the questions, the data selected and the computing resources, and it held the right to redact the report. The site says the published investigation covered only 24% of the timeline OpenAI itself had presented. Evaluators see only what a lab authorizes, it argues, and nothing requires them to disclose those limits in their findings.

That question is already in front of regulators. The Federal Trade Commission is investigating OpenAI, Anthropic and METR over AI agents, as Tech AI Wire reported on October 1.

Legal protection for the fired researchers is unclear. Progressive Robot notes that California's frontier AI law, SB 53, protects disclosures to state authorities or through internal channels. Whether sharing with a private safety group is protected has not been tested in court.

What this means for developers

Teams building on OpenAI's models get their safety assurances from two places: the company's own reports and the outside evaluations it commissions. If Business Model Analyst's account of the METR investigation is right, the second channel runs on OpenAI's terms. When you read an outside report on an OpenAI model, check what scope the lab set before treating the findings as independent.

If you run a safety or security team of your own, settle the policy question in advance. Decide now what staff may share with external researchers, through which channel, and who approves it. OpenAI's statement frames the firings as a procedure violation, not a disagreement about the content. That is where the fight will land at most companies.

Watch three things. First, whether OpenAI names the recipient organization or discloses what was shared. Second, whether METR or Redwood Research describe the access they had during their OpenAI work. Third, how the FTC's inquiry treats the relationship between labs and their evaluators, since that is the mechanism this case puts in doubt.