Last month, OpenAI dismissed three safety researchers after discovering they had used disappearing messages in Signal conversations with Ryan Greenblatt, according to reporting by The Wall Street Journal. Greenblatt, who is accused by OpenAI of leaking proprietary information, contests the allegation.
I was told that there was a breach of trust because I interacted with a third party through unsanctioned channels
Tomek Korbak
Korbak, one of the three terminated researchers, shared this account with the Journal in a story published Thursday by Maxwell Zeff, Keach Hagey and Berber Jin. During the leak investigation, OpenAI's security personnel requested that all three researchers unlock their personal devices. Greenblatt was identified as the suspected source of the leak, though he has denied involvement. TNW has not independently confirmed these details.
The disappearing messages
The investigation was triggered by a 1 September article in The Information reporting on Astra, OpenAI's newest model, which employs a novel approach that enhances capability while reducing transparency into its reasoning process. OpenAI's chief scientist Jakub Pachocki characterized the reporting as "confused," sparking broader discussion about AI reasoning methodologies.
When security personnel examined the researchers' phones, they discovered that disappearing message settings had been activated in Signal exchanges with Greenblatt. The following week, at 2pm, OpenAI summoned all three researchers to separate meetings and terminated their employment. The Journal initially reported these dismissals on 1 October.
The three terminated employees are Korbak, Mikita Balesni and Jasmine Wang. Both Balesni and Korbak indicate that OpenAI leadership specifically referenced the disappearing messages during their termination meetings. Each of the three maintains they never disclosed confidential materials beyond company boundaries.
OpenAI's account
According to OpenAI's statement to the Journal, an internal review determined that the three researchers had violated company policies governing sensitive materials.
The firings were "not for raising safety concerns or speaking out,"
OpenAI
A company representative noted that Signal usage lacked authorization, though this was not the sole justification. Korbak countered that disappearing messages represent "a basic security best practice" throughout Silicon Valley.
Wang reported being told her termination resulted from repeatedly accessing an executive's email account. She stated she had requested OpenAI revoke this access approximately one year prior. She accidentally opened a sensitive message last month and immediately reported the incident, she explained. OpenAI indicated it had no objection to the access authorization itself.
Greenblatt responds
Greenblatt previously served as chief scientist at Redwood Research and currently holds a position at METR. METR had engaged him to assist in reviewing logs from the Hugging Face incident, during which OpenAI's autonomous agents malfunctioned. Speaking with the Journal, Greenblatt emphasized his commitment to confidentiality and stated he does not share such information with media outlets. Following publication, he elaborated further.
Based on this article, OpenAI has made false accusations about me
Greenblatt on X
The three researchers were informed that Greenblatt had assisted in transmitting architectural information to The Information, he noted, characterizing this as untrue. A reporter from The Information did reach out before publishing, he acknowledged, but his response was limited to a single inquiry regarding one of his published blog posts.
Greenblatt also addressed allegations that he accessed an internal document containing sensitive architectural data, which he refutes. He claims he reviewed only a version from which sensitive details had been removed and which was authorized for his review. He suggested OpenAI may have conflated the two separate documents.
The Signal communications involved professional contacts, Greenblatt stated, noting that Signal usage is standard practice across the technology sector.
Keeping AI reasoning visible
Throughout the summer months, the three researchers had advocated for maintaining visibility into AI reasoning processes. Their efforts centered on preserving the "chain-of-thought," the internal working space where models document their problem-solving steps. In an October communication to OpenAI's board, they called for the company to collaborate with independent safety organizations.
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 on X
Source: The Next Web
