Three researchers dismissed from OpenAI for allegedly mishandling information have published an open letter saying that was not the reason. "I believe we were fired for prioritising safety over the near-term interests of OpenAI as a corporation," wrote Mikita Balesni on X, alongside a joint letter with Tomek Korbak and Jasmine Wang. All three say they worked in safety or alignment roles until recently.
The company's position has not moved. On 2 October a spokesperson told the BBC its investigation "confirmed that these individuals mishandled sensitive information outside established company procedures", and on Friday research leaders repeated it: "We want to be very clear that these decisions were not about raising safety concerns or speaking out." The note said an internal investigation had uncovered "a significant breach of trust beyond what's outlined in the letter they published".
The firings themselves were reported here in the piece on the dismissals. This is the other side of it.
What the researchers say they were working on
Alignment is the practice of building human values into a system so it stays pointed at what people actually want. Korbak was specific about his concern, and it is the most technically interesting part of the letter.
"For months, I'd been raising safety concerns that we're losing the ability to monitor what AI agents think, one of our best tools for catching when they misbehave. I believe that was why I was fired."
That claim is worth unpacking, because it is not vague. Current systems can be made to show their working — a chain of reasoning a human can read before the model acts. It is an imperfect window, but it is a window. As models are trained to be more efficient, that visible reasoning can become compressed, abbreviated or internal, and the window closes. You are then left judging an agent only by what it did, after it did it.
Wang's concern was about the precedent rather than the technology: "We were not the first to be pushed out of OpenAI under suspicious circumstances. Unless the employees take a stand now against this kind of manoeuvre, I am concerned we will not be the last."
Why an outsider cannot settle this
Two accounts are irreconcilable and both are plausible. Someone can mishandle information and have raised safety concerns; a company can be telling the truth about a breach of trust and still be glad to lose an internal critic. Nothing published so far decides it.
What is observable is the effect the former staff describe: colleagues "afraid to speak and operate in ways that, until last week, were an integral part of working at OpenAI". Whether or not the firings were retaliatory, a workforce that believes they might be behaves the same way — and that chilling effect does not require anyone to have acted in bad faith.
OpenAI said it is finalising contracts with third-party safety assessors and will announce details in coming weeks. That is the substantive response to the underlying problem, because external assessment does not depend on an employee being willing to take the risk.
Why this reaches a buyer in Dhaka
Internal dissent at a San Francisco company looks remote from a procurement decision in Bangladesh. It is not, for one reason.
Almost everything a buyer here knows about a frontier model's safety comes from the company that sells it. There is no Bangladeshi regulator testing these systems, no local audit requirement, and no realistic way for a bank or a hospital to evaluate a model's behaviour at depth. The internal safety team is, in practice, part of the assurance being purchased.
So when that team is in open dispute with its employer, the thing a buyer should notice is not who is right. It is that the assurance rests on a process nobody outside can see — which is the argument for insisting on independent assessment, and for reading the eventual third-party contracts closely when they are announced.
The pattern across the sector is in the model OpenAI pulled over safety and the taskforce now coordinating US policy.
