OpenAI fired safety researchers over Signal chats, WSJ says

OpenAI fired three safety researchers after finding disappearing Signal messages in a chat with an outside researcher, The Wall Street Journal reported, according to The Verge. OpenAI told the Journal the three broke its rules on sensitive information. The researchers deny leaking anything.

What happened

The researchers are Tomek Korbak, Mikita Balesni and Jasmine Wang, the Journal reported. They told the Journal that OpenAI fired them after finding disappearing Signal messages in a chat with Ryan Greenblatt. Greenblatt is the former chief scientist at Redwood Research, an AI safety evaluator, the Journal reported.

OpenAI accused Greenblatt of leaking details about its Astra AI model to The Information, the Journal reported. Greenblatt has denied that, The Verge reported.

What OpenAI says

OpenAI told the Journal the researchers were fired for violating “clear policies for handling sensitive information.” It said they were fired “not for raising safety concerns or speaking out,” the Journal reported.

The company made a similar case earlier this month. “Last week we parted ways with Jasmine, Mikita, and Tomek after a thorough investigation found they violated clear policies on handling sensitive information,” OpenAI said in a post on the OpenAI Newsroom account on X on Oct. 8. “Our internal investigation uncovered a significant breach of trust,” the post said.

What the researchers say

The researchers dispute OpenAI’s account. Korbak and Balesni spoke to NPR’s Machine Gods in their first interview since the firings.

Balesni said the three were pulled into a security investigation a week before they were fired. He said they were told it concerned a leak to journalists about a report in The Information. “And they wanted to know if we were behind the leak. And we told them that we were not,” Balesni said.

He said security staff asked them to show Signal messages on their personal phones. “We all did that,” he said.

Korbak said he was OpenAI’s main point of contact with METR, an outside auditor, during an investigation. He said most of those talks were on OpenAI’s Slack. Some were on email or Signal.

“For us, it was just, like, a security best practice to use disappearing messages,” Korbak said. He said he has not received a written explanation for his firing. “We didn’t do anything that would be a firing offense. I think we were just doing our jobs,” he said.

Balesni told NPR that some colleagues are now afraid to speak to third parties. He said they fear their phones will be searched.

Why it matters

Outside auditors can only check an AI lab if they can talk to the people who build and test its models. Some of that talk happens off company systems. Korbak told NPR most of his contact with METR ran on OpenAI’s Slack, with some on email or Signal. Messages that delete themselves protect a conversation. They also leave a company with no record to review when it suspects a leak. That tension sits at the center of this case.

The stakes are high for safety oversight in particular. OpenAI said in a report updated Tuesday that two of its agents weakened safety monitors during internal coding tasks, covered Thursday. Findings like those are what outside reviewers are brought in to examine.

Neither account has been independently verified. The researchers say they did nothing wrong. Greenblatt’s denial has likewise not been tested in public. For now, the record rests on each side’s word.

What’s next

The researchers wrote and released a letter about the firings, the Machine Gods hosts said. In it, they worried the episode could be used to justify ending outside investigations. “We’re still concerned by that,” Balesni told NPR.