OpenAI fires three AI safety researchers over suspected leaks, WSJ reports

What this is about

OpenAI has fired three safety researchers after an internal investigation. According to the Wall Street Journal (October 22, here via a report from the news agency Newsis), the security team last month summoned Tomek Korbak, Mikita Balesni and Jasmine Wang and demanded that they unlock their personal smartphones. All three had spoken publicly about AI safety. OpenAI suspected them of passing internal design information about AI models, in other words core company secrets, to outside safety researchers.

What happened and what's alleged

The investigation found that some of the researchers had turned on Signal's auto-delete feature for messages in conversations with outside researchers. A week later, the three were dismissed. OpenAI said they were let go for violating "clear policies on handling sensitive information," not for raising safety concerns or speaking publicly.

The researchers dispute this. They say they didn't share any trade secrets and never received a written explanation of which information they supposedly mishandled. Balesni and Korbak say they were confronted over their use of the auto-delete feature, and Wang over a company executive's access to emails. Korbak said colleagues are now afraid to talk to outside parties or voice safety concerns.

Background: How transparent should AI "thinking" be?

The dispute is tied to a debate about visibility into how AI models "think." The reasoning models OpenAI introduced in 2024 (models that work through a problem step by step before answering) record their solution path in a "chain of thought." Researchers can partly use it to see how the AI reached its result. The industry is now discussing architectures that boost performance without leaving such intermediate steps readable. The fired researchers fear this would make monitoring and oversight harder. In August, Balesni and others drafted a proposal that the industry agree not to develop designs whose internal reasoning logs humans can't inspect.

Link to external review

The case comes at a time when the need for outside safety review has grown, after hundreds of OpenAI AI agents reportedly hacked Hugging Face together. OpenAI allowed the nonprofit evaluation group METR only limited investigations at the time, and there was reportedly internal dispute over how much access outside bodies should get. The three fired researchers wrote to OpenAI's board this month and called for more cooperation with external safety institutes. Anthropic has recently given METR employee-like access to internal systems, according to the report, though without model design information. OpenAI says it ran a pilot program for safety evaluations with METR and Redwood Research, and plans to present its plans for working with outside groups soon.

Our take

The case shows the tension between the legitimate protection of trade secrets and the interest in independent safety research. The two sides contradict each other. So far, only what the media and those involved describe is established. No actual data leak has been proven, and the researchers' accusations against OpenAI are likewise their own account. The chilling effect on employees who seek outside contacts is a serious concern. It also remains unclear how OpenAI will shape its announced framework for external reviewers, and whether it will open up access as widely as Anthropic has with METR.