IT之家 News on October 25: Three OpenAI safety researchers who were fired over alleged leaks have given their side of the story. Tomek Korbak, Mikita Balesni, and Jasmine Wang denied leaking confidential information, questioned the official reasons given for their dismissal, and speculated that the matter was connected to their continued advocacy for greater transparency in AI systems.
After an internal investigation, OpenAI determined that the three individuals had violated regulations on the handling of sensitive information and dismissed them. Before being fired, all three had publicly expressed views on AI risks.
IT之家 noticed that, in an interview with The Wall Street Journal, several researchers refuted the official account. OpenAI told the newspaper that the dismissal had nothing to do with the safety concerns they had raised. But three researchers said the company never informed them in writing exactly what information had been mishandled. They also worry that this incident will make independent safety auditors reluctant to work with AI companies in the future.
The Wall Street Journal's report reconstructed the origins of the conflict. OpenAI's safety team suspected that the three individuals had leaked details of the company's internal model architecture to an external safety researcher, and demanded that they unlock their personal phones. Investigators found Signal chat records with disappearing messages enabled on the phones. A week later, all three were fired simultaneously. OpenAI said Signal is not an approved communication tool for the company, but emphasized that this was not the reason for the dismissals. Kolbeck countered that disappearing messages are a routine security practice in Silicon Valley.
Point of Contention: Whether to Allow Verification of AI's Reasoning Steps
At the heart of this dispute is a far-reaching technical question: whether AI companies should allow outsiders to view a model's thinking process. Reasoning models record intermediate steps of problem-solving 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, Baranes, Korbak, and Jasmine Wang, together with leading researchers from multiple AI labs, called for this verifiable capability to be continuously preserved. This summer, the issue became more urgent: several AI companies began internal discussions about a new class of models — ones that think for longer but record fewer intermediate steps. Baranes raised objections within her company and forwarded a related proposal from Ryan Greenblatt, who was then chief scientist at Redwood Research.
Kolbakh also maintains frequent communication with Greenblatt. The two collaborate via Slack and Signal, jointly analyzing a high-profile incident: the OpenAI agent attacking Hugging Face incident. Greenblatt was commissioned by METR to conduct this independent investigation.
A report about Astra triggered an internal investigation
On September 1, The Information published a report stating that OpenAI's new model Astra is more capable but exposes less of its reasoning process. This collaboration thus became the trigger for trouble. At the time, OpenAI chief scientist Jakub Pachocki commented on X that the report's logic was muddled, but he himself had previously warned that the entire industry is racing to develop models that are hard to monitor.
The Wall Street Journal reported that weeks later, OpenAI launched an internal leak investigation. Balesni said investigators determined Greenblatt was the recipient of the leaked information and sought to obtain the Signal chat records between them. Greenblatt denied passing confidential information to reporters.
The allegations against Wang Moli were established separately: OpenAI said she repeatedly accessed a senior executive's mailbox. Her explanation was that she had applied to revoke this permission nearly a year earlier; last month she merely opened a sensitive email by mistake, did not forward or disseminate its contents, and reported the error immediately. OpenAI stated that the permission itself was not the issue.
Kolbath believes the negative impact of the incident extends far beyond the three of them. He said colleagues are now afraid to communicate with outside parties, and afraid to raise safety concerns internally. The three researchers have written to OpenAI's board, calling for stronger cooperation with external safety organizations.
Another newly released on-site audit report also raises questions about OpenAI's safety processes. Auditors from METR and the Sequoia Research Institute found numerous security vulnerabilities in OpenAI's infrastructure during their verification, but OpenAI redacted nearly half of the content of the final report.
OpenAI has also released a new security architecture report; this architecture is disconnected from the internet when running. The company said the relevant vulnerabilities have been fixed and model training resumed in early October (month 10), and it also stated that multiple new external partnerships will be announced soon.