Fired OpenAI safety researchers Tomek Korbak, Mikita Balesni and Jasmine Wang say they were never told in writing what confidential information they allegedly mishandled. All three were dismissed at the same time after an internal investigation into handling of sensitive information, and each had spoken publicly about AI risks before the dismissal. The dispute matters because it could discourage independent safety auditors from working with AI companies.
Signal chats and Astra report behind dismissals
According to Wall Street Journal reporting, OpenAI security staff suspected the three of sharing details about internal model architecture with an external safety researcher. Investigators asked them to unlock personal phones, where they found Signal chats with disappearing messages enabled. OpenAI said Signal was not an approved communication tool but insisted it was not the reason for the terminations, while Korbak described disappearing messages as routine security practice across Silicon Valley.
The investigation followed a September 1 report by The Information that OpenAI's new model Astra was more capable but revealed less of its reasoning process. Chief scientist Jakub Pachocki dismissed that report as muddled on X, although he had previously warned of an industry-wide race toward models that can no longer be monitored. A few weeks later OpenAI opened the internal leak investigation, with investigators naming Ryan Greenblatt, then chief scientist at Redwood Research, as the suspected recipient. Greenblatt denies passing confidential information to journalists.
At the center stands the chain of thought, the internal scratchpad where reasoning models record intermediate steps of problem solving. Companies and safety researchers use these traces to understand how models reach decisions. In a 2025 paper, Balesni, Korbak and Wang joined researchers from several AI labs in calling for continued access to such traces. The issue sharpened over the summer amid discussion of models that think longer but write down fewer intermediate steps, with Balesni internally sharing a related proposal from Greenblatt.
What the conflict means for external AI audits
For companies that rely on external review of AI systems, the case changes the terms of cooperation with model providers. Independent auditors and enterprise buyers depend on access to reasoning traces, audit reports and incident analysis to assess risk before deployment. Korbak, Balesni and Wang worked with Greenblatt on analysis of an incident involving OpenAI agents and Hugging Face, for which Greenblatt had been hired by METR to conduct an independent investigation. If such contacts trigger leak inquiries, large customers may face narrower audit scope and slower incident reviews, while smaller firms lose a shared source of safety expertise.
Several facts remain unresolved and require direct verification with a vendor. Wang faced a separate accusation of repeatedly accessing an executive inbox, which she says she had asked to revoke nearly a year earlier, adding that she accidentally opened a sensitive email, did not share it and reported the incident at once. OpenAI says the access itself was not the issue. Separately, staff from METR and Redwood Research found numerous security gaps during an on-site audit, but OpenAI redacted nearly half the final report, then published its own report on a new security architecture without internet access and said it fixed the issues.
The marker to watch is whether OpenAI announces several new external partnerships it has promised and what access those partners receive to reasoning traces and unredacted audit findings. Training reportedly resumed in early October after the fixes. If new agreements restore detailed trace access and independent publication rights, external oversight keeps working tools; if access stays limited, monitoring of advanced reasoning models becomes harder to sustain.
