Three fired OpenAI safety researchers delivered a letter to the company's board and safety committees on October 7, 2026, with excerpts reported by The Wall Street Journal and other outlets on October 8. The three — Jasmine Wang, Tomek Korbak and Mikita Balesni — all worked on safety and alignment teams. OpenAI parted ways with them on October 1, citing violations of its policies on handling sensitive information, but did not name them publicly or identify the outside organization involved. Their letter does two things at once: it defends their conduct, and it puts a specific technical demand to the company.
Laying out each side's account
In the letter and in public posts, the three say their outside contacts fell within their job mandates. Korbak says he served as the technical contact for the evaluation organization METR during OpenAI's investigation of the Hugging Face incident, while the company was still developing procedures for an unprecedented inquiry. Balesni says the cross-company monitorability work he coordinated was flagged to his reporting line in advance, and that he removed sensitive details before sharing materials. Wang, in an October 8 thread, says she mistakenly opened a sensitive email from an executive because of inbox access granted for recruiting that she had asked IT to revoke, without the request being completed; she says she alerted the executive within minutes and asked again for removal. The three also deny being the source of a leak about less-monitorable architectures. These are the researchers' own accounts; the full letter has not been published and the details cannot be independently verified.
The real demand: keep the chain of thought monitorable
The letter's technical ask is precise: the industry does not yet know how to safely develop and deploy models it cannot monitor, and frontier companies should not pursue developments that further weaken chain-of-thought monitorability. It also urges OpenAI to work with third-party auditors and support an open, transparent safety ecosystem, and says the firings are chilling those who remain. An OpenAI spokesperson said the dismissals were not about raising safety concerns or speaking out, and shared part of an internal memo sent October 8 by a research leader that strongly agreed with the letter's recommendations. Notably, the system card for GPT-6 Astra already acknowledges that its chain of thought is harder to monitor than its predecessor's — exactly the direction the three are warning about.
When the rules are unwritten, secrecy and collaboration collide
The part other labs should study is not who is right, but where the boundary sits. Safety research inherently requires exchanging information with outside evaluators, while confidentiality duties say nothing may leave the building. If the line between the two is drawn only by after-the-fact investigations, every researcher doing legitimate collaboration will first ask whether it could become evidence in the next inquiry. The three are asking for the rules on outside work to be written down. If the company genuinely agrees with the letter's technical advice, the practical response is to publish what is allowed, what is not, and who approves it — and to state its third-party audit arrangements. Otherwise, what erodes first may not be chain-of-thought monitoring, but employees' willingness to raise concerns at all.
