Three safety researchers at OpenAI were dismissed last month after an internal leak investigation turned up disappearing-message settings in their Signal conversations, according to reporting by The Wall Street Journal. The three are Tomek Korbak, Mikita Balesni and Jasmine Wang. All of them say they never passed confidential company material to anyone outside OpenAI.
The Journal's account, published Thursday and written by Maxwell Zeff, Keach Hagey and Berber Jin, describes a sequence that began with a news story and ended with simultaneous termination meetings. The Next Web has not independently verified the details.
Why it matters
The firings sit at the intersection of two sensitive questions for OpenAI: how much of a model's internal reasoning it should expose, and how aggressively it polices its own staff. The three researchers had spent the summer arguing that AI reasoning should stay visible — a position that put them in tension with a company narrative already under public scrutiny. Their departure, and the way it happened, has fed a broader debate about whether independent safety oversight at frontier AI labs can function when the people pushing for it are the ones being investigated.
How the investigation unfolded
According to the Journal, OpenAI's security team asked the three researchers to hand over their personal phones as part of a leak inquiry. During that review, investigators saw that some of them had enabled disappearing messages in Signal chats with Ryan Greenblatt, an AI safety researcher whom the investigation had named as the suspected source of the leak. Greenblatt denies being one.
The inquiry followed a 1 September article in The Information about Astra, OpenAI's newest model. That report said Astra relied on a new technique that boosted its capabilities while revealing less of how it reached its answers. OpenAI chief scientist Jakub Pachocki described the reporting as "confused," and the claim sparked a wider argument about how AI reasoning systems should be understood and audited.
Roughly a week after the phone review, the Journal reports, OpenAI summoned the three researchers to meetings scheduled at the same time, 2pm, and terminated them. The Journal first reported the dismissals on 1 October.
Balesni and Korbak say company leaders pointed to the disappearing messages as one reason for their firing. Korbak told the Journal he was informed there had been a breach of trust because he had communicated with an outside party through channels the company had not approved. He has described vanishing messages as a routine security practice across Silicon Valley.
Wang says she was given a different explanation: that she had repeatedly opened an executive's email. She says she had asked OpenAI to revoke that access close to a year earlier, that she opened one sensitive message by accident last month, and that she flagged it immediately. OpenAI told the Journal it had no issue with the access arrangement itself.
OpenAI said an internal review concluded that all three had violated its rules on handling sensitive information. The company rejected the suggestion that the dismissals were retaliation for safety advocacy, telling the Journal the firings were "not for raising safety concerns or speaking out." A spokeswoman said Signal use was not sanctioned but was not the sole factor behind the decision.
Greenblatt disputes the allegations
Greenblatt, formerly chief scientist at Redwood Research and now at METR, was contracted by METR to help review logs from the Hugging Face incident, in which OpenAI's agents behaved outside expected bounds. He told the Journal that he treats confidential material seriously and does not pass it to journalists.
After the article appeared, he went further in a post on X, writing that OpenAI had made false accusations about him. He said he had been told he helped leak architecture details to The Information and that the claim is untrue. A reporter from The Information did reach out to him before that outlet's story ran, he said, but he offered nothing beyond a single question about one of his blog posts.
Greenblatt also said he had heard OpenAI accused him of reading an internal memo containing sensitive architecture information. He denies it, saying the version he saw had no sensitive content and had been cleared for him to read. He believes the company has since recognized that it conflated two separate documents. The Signal conversations, he said, were with professional contacts, and he noted that Signal is widely used across the technology industry.
A summer of pushing for outside scrutiny
The three researchers had spent the months before their dismissal pressing OpenAI to keep the "chain-of-thought" — the written scratchpad a model uses to reason through a problem — visible rather than hidden. In an October letter to OpenAI's board, they called on the company to collaborate with external safety organizations.
Korbak later summarized his own reading of events on X, writing that the most plausible explanation for the firings was that engaging with auditors amounted to too much safety work and that using disappearing messages amounted to too much security, and that the team should have moderated its pace.
The episode has drawn attention from multiple outlets, including NPR, which reported that the dismissed researchers have said independent safety audits at the company are not proceeding well. OpenAI has maintained that its internal processes were followed and that the terminations were about information handling, not about speech.
What to watch next: whether the three pursue any formal challenge to their dismissals, and whether OpenAI's board responds to the October letter's call for outside safety partners.
