“I was told that there was a breach of trust because I interacted with a third party through unsanctioned channels,” Tomek Korbak said.
Korbak is one of three OpenAI safety researchers fired last month. He spoke to The Wall Street Journal for a story by Maxwell Zeff, Keach Hagey and Berber Jin, published on Thursday.
OpenAI’s security team asked the three to open up their personal phones during a leak investigation, the Journal reported. Investigators named AI safety researcher Ryan Greenblatt as the suspected leaker. He denies it. TNW has not independently verified the account.
The disappearing messages
The leak inquiry followed a 1 September story in The Information. It said Astra, OpenAI’s latest model, used a new technique that made it more powerful but showed less of its thinking. OpenAI chief scientist Jakub Pachocki called the reporting “confused”, and the claim set off a wider debate over AI reasoning.
On the phones, security officers saw that some of the researchers had turned on disappearing messages in their Signal chats with Greenblatt. A week later, OpenAI called the three into meetings at the same time, 2pm, and fired them. The Journal first reported the firings on 1 October.
The three are Korbak, Mikita Balesni and Jasmine Wang. Balesni and Korbak say OpenAI leaders cited the disappearing messages as a reason. All three say they never shared confidential information outside the company.
OpenAI’s account
OpenAI told the Journal that an internal investigation found the three had broken its policies on sensitive information.
The firings were “not for raising safety concerns or speaking out,” the company said.
A spokeswoman said using Signal was not approved, but it was not the only reason. Korbak called disappearing messages a “basic security best practice” in Silicon Valley.
Wang said she was told she was fired for repeatedly opening an executive’s email. She said she had asked OpenAI to remove that access nearly a year earlier. She opened one sensitive email by mistake last month and reported it at once, she said. OpenAI said it had no problem with the access itself.
Greenblatt responds
Greenblatt was chief scientist at Redwood Research and now works at METR. METR had contracted him to help examine logs from the Hugging Face incident, when OpenAI’s agents went rogue. He told the Journal that he takes confidentiality seriously and never shares such information with reporters. After the story ran, he went further.
“Based on this article, OpenAI has made false accusations about me,” Greenblatt wrote on X.
The three researchers were told he had helped leak architecture details to The Information, he wrote, and that is false. A journalist from The Information did contact him before its story ran, he said, but he did not comment beyond one request about a blog post of his.
He wrote that he had heard OpenAI also accused him of reading an internal memo with sensitive architecture details, which he denies. He said he saw only a version without sensitive information that was approved for sharing with him. He believes OpenAI now knows it mixed up the two documents.
The Signal chats were with professional contacts, Greenblatt said, and Signal use is common in tech.
Keeping AI reasoning visible
The three researchers had spent the summer pushing to keep AI reasoning visible. Their focus was the “chain-of-thought”, the written scratchpad a model uses to work through a problem. In an October letter to OpenAI’s board, they urged the company to work with outside safety groups.
“Best explanation I have for our firing: talking to auditors was too much safety, disappearing messages were too much security, and we should have paced ourselves,” Korbak wrote on X.