OpenAI fired three AI safety researchers after asking them to show personal phones during a leak investigation, The Wall Street Journal reported on Thursday. Security staff found disappearing messages turned on in Signal chats with researcher Ryan Greenblatt, named as the suspected leaker. The three were called to simultaneous meetings at 2pm a week later and dismissed. The case matters because it ties internal control of sensitive model information to a dispute over visible AI reasoning.
Phone checks and the Astra leak inquiry
Reporters Maxwell Zeff, Keach Hagey and Berber Jin identified the dismissed staff as Tomek Korbak, Mikita Balesni and Jasmine Wang, with the firings first reported on 1 October. The inquiry followed a 1 September story in The Information that said Astra, OpenAI's latest model, used a new technique that increased power while showing less of its thinking. OpenAI chief scientist Jakub Pachocki called that reporting confused, and the claim started a wider debate over AI reasoning. Investigators treated Greenblatt as the suspected source, which he denies, and TNW said it had not independently verified the account.
Korbak said he was told there was a breach of trust because he interacted with a third party through unsanctioned channels. Balesni and Korbak said managers cited the disappearing messages as a reason, while all three said they never shared confidential information outside the company. OpenAI told the Journal that an internal investigation found violations of policies on sensitive information. A spokeswoman said Signal use was not approved but was not the only reason, adding the firings were not for raising safety concerns or speaking out.
Wang said she was told a different operational reason: repeatedly opening an executive's email. She said she had asked OpenAI to remove that access nearly a year earlier, opened one sensitive email by mistake last month and reported it at once. OpenAI said it had no problem with the access itself. Korbak described disappearing messages as a basic security best practice in Silicon Valley. Greenblatt said the Signal chats were with professional contacts and that Signal use is common in tech, while stressing that he takes confidentiality seriously.
What the dismissals mean for AI oversight
For companies that rely on external model evaluations, the episode shows how access rules and messaging tools can become part of vendor risk. Greenblatt was chief scientist at Redwood Research and now works at METR, which had contracted him to help examine logs from the Hugging Face incident when OpenAI agents went rogue. A small firm using such evaluations may face delays if communication channels are restricted, while a large enterprise may need clearer rules on approved messengers, retention, and who may view internal memos. Approved sharing versions and logging procedures become practical controls, not paperwork.
Several points remain contested and deserve checks before drawing broad conclusions. Greenblatt wrote on X that OpenAI made false accusations about him, denying that he helped leak architecture details to The Information. He said a journalist contacted him before publication but he did not comment beyond one request about his blog post. He also denied reading an internal memo with sensitive architecture details, saying he saw only a version approved for sharing and believes OpenAI mixed up two documents. This news by itself does not prove a leak occurred or clarify which policy clauses were breached.
The marker to watch is whether OpenAI clarifies the document mix-up described by Greenblatt and publishes clearer rules on reasoning transparency. The three researchers spent the summer defending the chain-of-thought, the written scratchpad a model uses to work through a problem, and urged board cooperation with outside safety groups in an October letter. If Astra documentation or audit access expands, outside review will get easier; if controls tighten without explanation, evaluation work will slow.
