OpenAI fired three safety researchers — Tomek Korbak, Mikita Balesni and Jasmine Wang — on October 1, after an internal leak investigation found that some of them had used disappearing messages on Signal in chats with an outside researcher, The Wall Street Journal reported on Thursday. OpenAI says its investigation confirmed the three mishandled sensitive information outside company procedures. The researchers dispute that account, and say they never shared OpenAI’s confidential information with anyone outside the company.
A Leak Investigation Over Astra’s Design
The investigation began with a September 1 story in The Information claiming that OpenAI’s latest model, Astra, employed a new technique that made it more powerful but revealed less of its thinking. Chief scientist Jakub Pachocki called the reporting “confused” on X, while raising his own concerns about a race toward models that cannot be monitored.
Within weeks, OpenAI announced internally that it was investigating the leak. Its global security team pulled researchers into meetings and asked them to unlock their personal phones. Investigators named Ryan Greenblatt — then chief scientist at Redwood Research, now at the evaluation nonprofit METR — as the suspected recipient, and found disappearing messages turned on for some of the researchers’ Signal chats with him.
A week later, the three were summoned to simultaneous 2 p.m. meetings and fired. Greenblatt denies he was the source of any leaks to The Information: “I take confidentiality very seriously and never share such information with reporters.”
What OpenAI Says, And What The Researchers Say
OpenAI told the Journal it fired the three after an internal investigation found they violated “clear policies for handling sensitive information,” and “not for raising safety concerns or speaking out.” A spokeswoman said using Signal was not approved, but was not the sole reason for the firings. The company has called the episode a “significant breach of trust.”
The researchers say they never received a written explanation of what sensitive information they allegedly mishandled. Korbak, 32, served as OpenAI’s technical point of contact for METR. “I was told that there was a breach of trust because I interacted with a third party through unsanctioned channels,” he said. “Talking to METR was my job.” He added that using disappearing messages on Signal is considered a “basic security best practice” throughout Silicon Valley.
Balesni, 27, said he was fired for “speaking too much to third-party safety organizations,” and that he shared a proposal with Greenblatt — stripped of company secrets — only after receiving authorization to speak with him. Wang, 28, said the company cited her access to an executive’s email, delegated to her for recruiting; when she accidentally opened a sensitive message, she reported it within minutes. The spokeswoman said the company did not have a problem with the access itself.
A Fight Over Keeping Models Monitorable
All three researchers built their reputations on chain-of-thought monitorability — the idea that a reasoning model’s written scratchpad is a rare window into how AI agents make decisions, and should be preserved. Before joining OpenAI in 2025, Balesni and Korbak co-led a paper making that case, with Wang a co-author alongside leaders from Anthropic, Google DeepMind and Meta.
This summer, as workers across the industry discussed new, more opaque model designs, Greenblatt drafted a proposal for companies to avoid the change. Balesni shared it in an OpenAI Slack channel that included Pachocki, and by August had co-written a similar internal proposal, which he says he discussed with a board member and a senior executive.
Days after their dismissals, the three wrote to OpenAI’s board urging the company to work with external safety organizations to help stave off the “risk that something truly catastrophic will happen,” and warning that the firings had created a “chilling” atmosphere for employees still working on alignment. Korbak says colleagues are now afraid to talk to third parties, or to raise safety concerns. The contrast inside the industry is sharp: Anthropic last week gave METR researchers employee-level access to its systems, while METR’s own red-teaming has found exploits in OpenAI’s agent monitoring. On Thursday, OpenAI said it had run a pilot embedding METR and Redwood researchers to assess its safety systems — though METR said OpenAI redacted some findings it could share.
Why It Matters
OpenAI’s handling of the three researchers matters because outside evaluation is one of the few independent checks on frontier models, and it runs on exactly the relationships this investigation punished. Korbak’s job was to be METR’s point of contact; Balesni says his contact with Greenblatt was authorized. If staff conclude that sanctioned conversations with evaluators can still end a career when a leak hunt starts, information will stop flowing to the outside groups that grade OpenAI’s safety work — just as the company asks to be trusted on models whose reasoning it may soon be unable to show anyone. The suspected leak was not about a product or a price, but about how a model thinks: architecture is now crown-jewel intellectual property.
Background
Chain-of-thought monitoring rose with the reasoning models that appeared from late 2024, which write out interim steps before answering and leave a legible trail safety teams can inspect. The fired researchers’ monitorability paper became a reference point for that debate, and METR sits at the centre of the external-audit ecosystem OpenAI itself helped create — even as the company has cycled through safety leadership since its 2024 governance crisis.
Source: The Wall Street Journal, AI Weekly
0 Comments