Fired OpenAI Safety Researchers Publish Open Letter

Three safety researchers dismissed by OpenAI reject the leak accusation in an open letter and press the company for three specific commitments on external review and model monitoring. Tomek Korbak, Jasmine Wang and Mikita Balesni all worked on safety and alignment at OpenAI. Their letter went out on X, and part of its audience is clearly the colleagues they left behind.

The argument: no lab makes AI safe by itself

t3n puts the letter's thesis in its headline — OpenAI cannot make AI safe alone. The researchers hold that work on the risks of large models breaks down without close contact with outside specialists. That contact, they argue, is what the dismissals now put at risk. They add a second point that is easy to miss: the ability to monitor current models is shrinking rather than keeping pace.

Two versions of why they left

According to heise online and t3n, OpenAI pointed to the sharing of internal information. The three dispute that version. Balesni describes receiving no written reason at all; in a closing meeting, he says, he was told the company no longer trusted him because he spoke frequently with organizations outside OpenAI. Those conversations, he counters, had been coordinated with management, research leadership and the board.

What the letter actually asks for

• External review: honor existing commitments to independent safety audits, and give outside specialists access comparable to what employees have.
• Monitorability: build out the tooling for watching frontier models instead of letting it fall behind what those models can do.
• Open exchange: let in-house safety researchers talk to the wider security ecosystem through clearly defined processes.

Why the tone is sharp

The authors describe a chilling effect: once staff see that outside contact can end a job, they speak more carefully inside the company too. On their reading, that costs OpenAI the external corrective that makes safety work credible in the first place. t3n places the letter alongside other recent episodes, including the dissolution of an internal team for AI crisis cases. That framing comes from the t3n report, not from the letter.

Open questions

Neither source we read carries a detailed OpenAI response to the individual demands. Neither gives the exact date the letter was posted on X; t3n dates only its own report, to October 9, 2026. Whether anyone beyond the three signed the letter is not stated in either piece. Nor do they document which contractual commitments the audit demand rests on — so that stays an open question here rather than a claim.