Three Fired OpenAI Safety Researchers Deny Leaking Secrets, Say It's Related to Promoting AI Transparency
Three safety researchers fired by OpenAI over suspected leaks—Tomek Korbak, Mikita Balesni, and Wang Moli—deny disclosing confidential information, question the official reasons for their dismissal, and speculate that the matter is related to their advocacy for greater transparency in AI systems. OpenAI says the dismissals are unrelated to the safety concerns raised by the three, but the researchers say the company never informed them in writing which information was improperly handled. At the core of the dispute is whether AI companies should allow outside parties to view the chain-of-thought records of reasoning models; the three called in a 2025 paper for this verifiable capability to be continuously preserved.
Why read this
AI score 64
After three researchers were fired, they publicly pushed back against the official account, an incident that has brought up the controversy over whether AI companies should preserve the verifiability of models' chains of thought.
Topics
Body · AI translation
Three Fired OpenAI Safety Researchers Deny Leaking Secrets, Suspect They Were Targeted for Promoting AI Transparency
English text is being prepared.
English text is being prepared.
After an internal investigation, OpenAI determined that the three had violated regulations on handling sensitive information and dismissed them. Before being fired, all three had publicly expressed views on AI risks.
English text is being prepared.
The Wall Street Journal report reconstructed how the conflict began. OpenAI's safety team suspected the three of leaking details of the company's internal model architecture to an outside safety researcher and demanded they unlock their personal phones. Investigators found Signal chat records with disappearing messages enabled on the phones. A week later, all three were fired at the same time. OpenAI said Signal is not a company-approved communication tool but stressed that this was not the reason for dismissal. Korbak countered that disappearing messages are a standard security practice in Silicon Valley.
At the heart of the dispute is a far-reaching technical question: should AI companies allow outsiders to view a model's thought process? Reasoning models record intermediate problem-solving steps in an internal "chain of thought" scratchpad, and companies and safety researchers rely on these records to understand the logic behind a model's judgments.
In a 2025 paper, Balesni, Korbak, and Wang, together with leading researchers from multiple AI labs, called for continuously preserving this verifiable capability. This summer, the issue became more urgent: several AI companies began internal discussions about a new class of models that think longer but record fewer intermediate steps. Balesni raised objections internally and forwarded a related proposal by Ryan Greenblatt, then chief scientist at Redwood Research.
Korbak also kept in frequent contact with Greenblatt. The two collaborated via Slack and Signal to jointly analyze a high-profile incident: the OpenAI agent attack on Hugging Face. Greenblatt was commissioned by METR to lead that independent investigation.
A report about Astra triggered an internal investigation
On September 1, The Information published a report saying that OpenAI's new model Astra was more capable but exposed less of its reasoning process. This collaboration then became the fuse for the trouble. At the time, OpenAI chief scientist Jakub Pachocki commented on X that the report was logically confused, though he himself had previously warned that the entire industry was racing to develop models that are hard to monitor.
The Wall Street Journal said that a few weeks later OpenAI launched an internal leak investigation. Balesni said investigators identified Greenblatt as the recipient of the leaked information and demanded access to their Signal chat records. Greenblatt denies passing confidential information to journalists.
The allegation against Jasmine Wang stands separately: OpenAI says she repeatedly accessed an executive's email inbox. Her explanation is that she had applied to have that permission revoked nearly a year earlier; last month she merely accidentally opened a sensitive email, did not forward or disseminate its contents, and reported the mistake immediately. OpenAI said the permission itself was not the issue.
Korbak believes the negative impact of the incident goes far beyond the three of them. He said colleagues are now afraid to communicate with outsiders and afraid to raise safety concerns internally. The three researchers have written to OpenAI's board of directors, calling for stronger cooperation with external safety organizations.
In addition, a just-released on-site audit report also raises questions about OpenAI's safety processes. Auditors from METR and Redwood Research found numerous security vulnerabilities in OpenAI's infrastructure during their review, but OpenAI redacted nearly half of the final report.
OpenAI also released a new security architecture report; this architecture disconnects from the internet when running. The company said the relevant vulnerabilities have been fixed, resumed model training in early October, and said it will soon announce several new external partnerships.