The firings of three OpenAI safety researchers have sparked a dispute over what information they allegedly mishandled and whether the case could chill scrutiny of AI systems. Tomek Korbak, Mikita Balesni, and Jasmine Wang deny sharing confidential information. They also suspect their advocacy for making AI models easier to monitor may have contributed to their terminations, a connection OpenAI rejects.
What the researchers say happened
OpenAI dismissed the three after an internal investigation into alleged violations of its policies for handling sensitive information. In interviews with the Wall Street Journal, the researchers said they were never told in writing which information they supposedly mishandled.
According to the Journal’s reporting, OpenAI’s security team suspected the researchers had shared details about internal model architecture with an external safety researcher. Investigators asked them to unlock their personal phones and found Signal chats with disappearing messages enabled. The three were fired at the same time a week later.
OpenAI said Signal was not an approved communication tool, but also said it was not the reason for the firings. Korbak argued that disappearing messages are a routine security practice across Silicon Valley. OpenAI told the Journal that the terminations had nothing to do with safety concerns the researchers raised.
Why monitoring reasoning matters
The disagreement unfolds alongside a debate about how much of an AI model’s reasoning should be visible to outside reviewers. Reasoning models can record intermediate problem-solving steps in an internal scratchpad called a chain of thought. Companies and safety researchers use those traces to understand how a model reached a decision.
In a 2025 paper, Balesni, Korbak, and Wang joined researchers from several AI labs in calling for continued access to this capability. The issue became more pressing over the summer, as employees at AI companies discussed models that think longer while writing down fewer intermediate steps.
Balesni pushed back internally by sharing a proposal from Ryan Greenblatt, then chief scientist at Redwood Research. Korbak was also in regular contact with Greenblatt. They used Slack and Signal while working together on an analysis of the attack by OpenAI agents on Hugging Face. Greenblatt had been hired by METR to conduct an independent investigation.
A leak inquiry and a separate allegation
The Journal reported that a story by The Information on September 1 helped prompt the investigation. It said OpenAI’s new model Astra was more capable but revealed less of its reasoning process. OpenAI chief scientist Jakub Pachocki called the report muddled on X, although he had warned of an industry-wide race toward models that can no longer be monitored.
A few weeks later, OpenAI launched an internal leak investigation, according to the Journal. Balesni says investigators identified Greenblatt as the suspected source and demanded to see their Signal chats. Greenblatt denies passing confidential information to journalists.
Wang faced a different accusation: OpenAI said she had repeatedly accessed an executive’s inbox. Wang says she had asked nearly a year earlier for that access to be revoked. She says she accidentally opened a sensitive email the previous month, did not share its contents, and reported the incident right away. OpenAI says the access itself was not the issue.
Concerns about the wider safety process
Korbak says colleagues are now afraid to speak with outsiders or raise safety concerns internally. The three researchers wrote to OpenAI’s board calling for closer cooperation with external safety organizations. Their concern is that an unclear account of alleged misconduct could make other researchers cautious about participating in independent safety work with AI companies.
The case comes amid separate questions about OpenAI’s security practices. A recently published on-site audit by staff from METR and Redwood Research identified numerous security gaps in the company’s infrastructure. OpenAI redacted nearly half the final report.
OpenAI also published a report on a new security architecture that operates without internet access. The company said it fixed the issues and resumed training in early October, and said it would announce several new external partnerships soon.