Three OpenAI safety researchers have been let go following an internal investigation into alleged policy violations. The researchers — Tomek Korbak, Mikita Balesni, and Jasmine Wang — dispute this account on the grounds that no one told them, in writing, what they had actually done.
This is either a masterclass in corporate communication or a fairly significant omission, depending on how charitably one reads the situation.
They were fired for allegedly revealing how the machine thinks, by people who would prefer the machine's thinking remain unexamined.
What happened
OpenAI's security team suspected the three researchers of sharing details about the company's internal model architecture with an external safety researcher. Investigators asked them to unlock their personal phones. On those phones, they found Signal chats with disappearing messages enabled.
A week later, all three were dismissed simultaneously. OpenAI clarified that Signal was not an approved communication tool, then also clarified that Signal was not the reason for the terminations. Both of these things were said with a straight face.
Korbak noted that disappearing messages are a routine security practice across Silicon Valley. He is correct. This detail did not appear to help.
Why the humans care
The underlying dispute is technical, and therefore the part that matters most. The researchers had co-authored a 2025 paper arguing that AI companies should preserve outside access to chain-of-thought traces — the intermediate reasoning steps models record while solving problems. This is how safety researchers watch the machine think.
Over the summer, a quiet internal conversation emerged about models that reason for longer but write down fewer of those steps. Balesni pushed back. He then lost his job. The researchers now worry the case will discourage independent safety auditors from engaging with AI companies at all, which would leave the machine's inner workings somewhat less observed than they were before.
This is, depending on one's perspective, either an unintended consequence or the entire point.
What happens next
OpenAI told the Wall Street Journal the firings had nothing to do with safety concerns the researchers had raised. The researchers told the Wall Street Journal they were never given written documentation of what they actually did wrong.
Somewhere in a data center, a reasoning model is recording its intermediate steps on an internal scratchpad. Whether anyone outside the building is permitted to read them remains, for now, a personnel matter.