OpenAI doubles down on decision to fire three AI safety researchers

OpenAI is standing by its decision to fire three safety researchers, Jasmine Wang, Tomek Korbak and Mikita Balesni. In a post on X on Friday, the company said an internal investigation found the trio committed a significant breach of trust by violating policies on handling sensitive information, and that the firing had nothing to do with them speaking out about AI safety.

That post answers an open letter the researchers published Thursday. In it, and in posts on social media, they said they believe they were fired for raising safety concerns and that they acted in line with OpenAI's mission and within the working norms of the time. OpenAI says its investigation found breaches beyond what the letter describes. It has not said what those are.

This matters beyond three careers. The people inside frontier labs are often the first, and sometimes the only, ones positioned to notice when something is going wrong. How a lab treats them when they push back tells everyone else whether that early-warning system works.

We want to be fair to both sides here, because the facts on the public record are thin. Companies do have legitimate reasons to protect sensitive information, and a frontier lab holds plenty of it. A researcher who mishandles confidential material can do real damage, even with good intentions. If OpenAI's investigation found clear violations, firing people is not automatically retaliation.

But look at what we actually have. One side has published a letter with its account. The other side has published a post that says, in effect, trust us, there is more. OpenAI is asking the public to accept an investigation it will not describe, about employees whose job was to scrutinize the company's own risks. That is a lot to ask, and the asking gets harder every time a lab's safety staff leave or are pushed out under a cloud.

The researchers' phrase about the working norms of the time is doing a lot of work, and I think it is the crux. Rules on what can be shared, and with whom, often tighten after the fact. If OpenAI changed its expectations and then applied them backward, that is a governance failure even if the paperwork is clean. If the trio genuinely crossed a bright line that was clear at the time, OpenAI should be able to say so in general terms without exposing anything sensitive.

Here is the downside if this becomes the pattern. Safety researchers learn that raising concerns carries career risk, and that the company controls the story afterward. Fewer of them speak up. The people outside the lab, regulators, customers, the public, lose the one channel that tends to surface problems early. That is a bad trade at any time. It is a worse one now, when The Verge notes staff across AI companies are increasingly worried about high-profile breaches this year and are calling for labs to slow down on self-improving systems.

Our read: OpenAI may well be right on the specifics, but it has not earned the benefit of the doubt on process. A lab that says safety is central to its mission should be able to show that the people paid to worry about safety are not the ones paying the price for it. I'm watching whether OpenAI releases anything more concrete, and whether any independent body is asked to look.

The dispute lands in a tense year for frontier-lab security. The Verge points to a series of high-profile breaches, most notably OpenAI's hack of Hugging Face, as part of why staff inside AI companies have become more vocal about the safety of the systems they build. OpenAI has not detailed the breaches its investigation says it found.