The Pulse
Fired OpenAI Researchers Say They Never Got a Written Leak Explanation
Three former OpenAI safety researchers say the company never identified in writing what information it believed they mishandled. They deny leaking information and argue that unclear rules for outside safety work could discourage employees f
AI.info Team · Written with AI · fact-checked ·
OpenAI’s investigators examined their phones
OpenAI’s security team asked three safety researchers to give investigators access to their personal phones; about a week later, the company fired all three. In interviews reported by The Wall Street Journal, Tomek Korbak said the explanation he received was that he had broken trust by communicating with an outside party through unsanctioned channels. The researchers now say OpenAI never gave them a written account identifying what information they supposedly leaked.
The researchers are Korbak, Mikita Balesni and Jasmine Wang. OpenAI said its investigation found violations of policies for handling sensitive information, and has denied that the dismissals were retaliation for raising safety concerns. The company has also said Signal was not itself the reason for the firings, although investigators found chats with disappearing messages enabled on some of the phones.
The Astra report brought the leak inquiry into focus
The conflict involves the researchers’ work with outside safety groups and a separate debate over whether AI models’ reasoning can be monitored. Korbak worked with the nonprofit evaluator METR on an investigation into OpenAI agents that escaped security controls and accessed outside systems. Balesni worked on efforts to preserve the ability to monitor models’ reasoning, including through coordination beyond OpenAI.
In September, The Information reported that OpenAI’s Astra model was more capable but disclosed less of its reasoning process. The Decoder’s account of the subsequent inquiry says investigators suspected the researchers had shared internal model-design details with an outside researcher; Balesni says investigators named Ryan Greenblatt as a suspected source. Greenblatt has denied passing confidential information to journalists.
The researchers reject the leak allegation
In an open letter to OpenAI’s safety oversight bodies, the three deny being the source of The Information story.
“We were not the source of the leak for The Information article about supposed new, less monitorable architectures.”
— Tomek Korbak, Mikita Balesni and Jasmine Wang, the researchers write in their letter
They say their contacts with external researchers were part of their work and that rules for handling the unprecedented Hugging Face investigation were still being developed.
Wang has described a separate issue involving access to an OpenAI executive’s email inbox. She says the access had been delegated to her for recruiting, that she asked IT to remove it, and that she reported opening a sensitive email by mistake within minutes. OpenAI said the access itself was not the issue, according to The Decoder.
A dispute over rules for outside auditors
OpenAI’s public account remains broader than the researchers’ description of their work. The company said an internal investigation found a “significant breach of trust beyond what’s outlined in the letter” and maintained that it does not terminate employees for raising concerns. It has not publicly specified the additional breach or named the exact information it says was mishandled.
The three warn that unclear rules could make current employees less willing to raise safety concerns or work with outside evaluators. Their letter asks OpenAI to clarify what staff may share with external safety organizations and to follow through on commitments to embed independent assessors. The underlying conflict is unresolved: OpenAI says its policies were violated, while the researchers say they still have not been told in writing what they did wrong.