OpenAI fired three of its safety researchers last month after an internal investigation concluded they had mishandled sensitive information, the company confirmed to The Wall Street Journal. The three, Tomek Korbak, Mikita Balesni and Jasmine Wang, had spent their final months at the company pushing OpenAI and its rivals to keep the inner workings of advanced AI models visible to outside auditors. OpenAI says they were dismissed for violating "clear policies for handling sensitive information," and "not for raising safety concerns or speaking out." All three dispute that account and say they never received a written explanation of what they supposedly leaked.
The firings grew out of a leak investigation. OpenAI's security team called the researchers into meetings and asked them to unlock their personal phones, suspecting they had passed details of the company's most closely guarded secret, the internal designs of its models, to an outside safety researcher. Investigators found some of the researchers had turned on disappearing messages in Signal chats with Ryan Greenblatt, then chief scientist at the nonprofit Redwood Research. OpenAI cited that use of disappearing messages as a reason for the terminations, though it said it was not the sole one. Wang said she was told she was fired for repeatedly accessing an executive's email, which she says happened once, by mistake, after she had asked a year earlier to have the access removed.
At the center of the dispute is a feature called the chain-of-thought. Reasoning models, introduced across the industry in late 2024, work through problems on an internal scratchpad, and that scratchpad offers researchers a rare window into how an AI system reaches its decisions. Balesni and Korbak co-led a 2025 paper, with co-authors from OpenAI, Anthropic, Google DeepMind and Meta, arguing that this visibility should be preserved. Their worry, they say, is that newer architectures designed to reason longer without writing down their steps would close that window. A September report by The Information claimed OpenAI's latest model, Astra, used a technique that made it more powerful while revealing less of its thinking; OpenAI's chief scientist Jakub Pachocki called the reporting "confused," then launched the leak inquiry that swept up the three.
The episode fits a wider pattern of labs wrestling with how much to show outsiders. Anthropic last week gave staff from the safety nonprofit METR "employee-like access" to its systems, but carved out model architectures as too risky to share. OpenAI said it ran a pilot embedding researchers from METR and Redwood to assess its safety systems, and that METR redacted some findings. The firings land as OpenAI chief executive Sam Altman warns publicly that advanced models will eventually escape any sandbox and that only alignment, not containment, can keep them in check, and as he floats governance schemes ranging from an international inspection body to a merged industry lab. A separate self-regulatory effort, the Standards Authority for Frontier AI, is being organized by OpenAI, Google and Anthropic.
What remains unresolved is whether internal dissent can survive the race it is meant to check. "We are hearing from our colleagues that they are afraid to talk to third parties now, and are afraid of raising safety concerns," Korbak said. OpenAI's own research leaders have said that preserving the monitorability of frontier models requires an industrywide commitment, including from OpenAI. The three fired researchers made the same argument in an October letter to the board, urging the company to work with external groups to head off the "risk that something truly catastrophic will happen." The company that agrees with them in principle has just dismissed the people who pressed the point hardest.