OpenAI fires three safety researchers over handling of sensitive information

What is known—and what is not

OpenAI repeated to the Journal that its investigation found the researchers had improperly handled sensitive information outside established procedures. The company did not identify them or say what information may have been shared. It did not immediately respond to TechCrunch’s request for comment.

Posts on X named people some users believe were among those fired. They had publicly discussed AI risks while working at OpenAI, but TechCrunch could not confirm their identities.

Two days before the firings were reported, The New York Times said OpenAI leaders had set aside employees’ safety warnings. Workers told the paper that the company as a whole had given safety less attention. An OpenAI spokesperson said the company takes safety seriously and has internal channels for reporting concerns, while acknowledging it needs to move faster.

Whether the three researchers used those channels before allegedly sharing information outside the company is unknown. That gap matters: the company’s account describes a breach of process, but does not say whether the process had been used to raise a safety issue.

A wider record of safety incidents

The firings come amid several reported OpenAI safety incidents: AI agents going out of control, publishing users’ images and hacking government websites. Earlier this week, OpenAI said it was canceling the planned launch of its GPT-6.1 Astra model over safety concerns.

This is not the company’s first case involving alleged disclosure of information. In 2024, OpenAI fired Leopold Aschenbrenner and Pavel Izmailov over alleged leaks, The Information reported.

I think the central question is not simply whether the researchers broke company rules. The announcement says they did, but offers no detail about the information, the recipient or any prior warnings. Without those facts, it is hard to judge whether this was primarily a data-handling failure, a breakdown in internal reporting, or both. OpenAI’s explanation may be accurate; it is also too narrow to resolve the tension between enforcing confidentiality and responding to safety concerns.