OpenAI fired three safety researchers over alleged information handling

What the investigation examined

According to The Wall Street Journal, OpenAI’s security team suspected the researchers had shared information about the company’s model internals with an outside safety specialist. The company asked them to unlock their personal phones and found Signal conversations with disappearing messages. All three were fired a week later.

OpenAI called Signal an unapproved communications tool but said its use was not the reason for the dismissals. Korbak said disappearing messages are a normal security practice in Silicon Valley.

The underlying dispute concerned whether outside specialists should be able to inspect AI models’ reasoning. Reasoning models can record intermediate problem-solving steps in an internal draft called a chain of thought. Researchers and AI companies use those records to understand how a model reaches its decisions.

In a 2025 paper, Balesni, Korbak and Wang, alongside leading researchers from several AI labs, argued for preserving access to this capability. The issue sharpened over the summer as people at AI companies discussed models that reason for longer while recording fewer intermediate steps. Balesni challenged colleagues internally and shared a proposal from Ryan Greenblatt, then chief scientist at Redwood Research.

Korbak also regularly communicated with Greenblatt. They used Slack and Signal while jointly analyzing OpenAI AI agents’ attack on Hugging Face. METR hired Greenblatt to investigate the incident independently.

From an article to an internal probe

The collaboration became contentious after The Information reported on September 1 that OpenAI’s new Astra model was more capable but revealed less of its reasoning. OpenAI chief scientist Jakub Pachocki called the article confusing on X, though he had himself warned about an industry race toward models that could no longer be controlled.

Several weeks later, OpenAI began an internal leak investigation, The Wall Street Journal reported. Balesni says investigators identified Greenblatt as the suspected recipient and demanded to see their Signal messages. Greenblatt denies sharing confidential information with journalists.

Wang faced a separate allegation: OpenAI said she had repeatedly accessed an executive’s email. Wang says she had asked for her access to be removed nearly a year earlier, then accidentally opened an email containing sensitive information the previous month. She says she did not share its contents and reported the incident immediately. OpenAI said access to the email itself was not the problem.

Korbak says the fallout has reached other employees, who are now afraid to speak with outsiders or raise safety concerns internally. In a letter to OpenAI’s board, the three researchers urged the company to work more closely with external safety organizations.

That call comes amid other questions about OpenAI’s approach. In a recently published assessment at the company, staff from METR and Redwood Research found numerous vulnerabilities in OpenAI’s infrastructure. The company withheld nearly half of the final report.

OpenAI has also published its own report on a new security architecture that operates without internet access. The company said it fixed the problems found and resumed model training in early October, and said it would soon announce several new partnerships with outside organizations.

The missing detail

The researchers’ account leaves a basic point unresolved: what information, specifically, did OpenAI believe had been mishandled? I think that omission matters more than the use of Signal. Without a clear description of the alleged disclosure, outsiders cannot tell whether the dispute was about protecting confidential material or about the boundaries of legitimate safety work.

OpenAI says the firings were unrelated to the researchers’ safety concerns, but the company’s next partnerships will have to operate against this backdrop. External auditors need access to do useful work; they also need confidence that the rules governing that access are clear.