AI Safety

OpenAI Fired 3 Safety Researchers; They Dispute Why

Oct 9, 2026·4 min read·By Rahish Saifi

Cover: 3 Fired OpenAI Safety Researchers Push Back — open letter vs company response — AI with Rahish

OpenAI fired three safety researchers last week (the company says it “parted ways” with them)—Jasmine Wang, Tomek Korbak, and Mikita Balesni—in a dispute that quickly spilled into public view. On October 8, 2026, the trio published an open letter titled “OpenAI cannot make AI safe on its own” (PDF), warning that messaging around their exits had left colleagues “afraid to speak.” On October 9, OpenAI posted a statement on X attributed to its research leaders, which Reuters and CNBC also reported, defending the firings as enforcement of clear policies on sensitive information, not retaliation for raising safety concerns.

The letter and the company response agree on one uncomfortable premise: frontier AI is not a normal product, and governance choices at OpenAI ripple across the wider safety ecosystem. They disagree sharply on why these three researchers left—and on what that means for internal culture and external oversight.

What the open letter says

The letter is addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. Its authors argue that “AI is not a normal technology, and OpenAI is not a normal company.” They deny being the source of the leak behind The Information’s article about “supposed new, less monitorable architectures”, and deny engaging external parties beyond what their jobs required.

They also place the dispute in recent operational context: the Hugging Face incident investigation, which they call “without precedent,” saying “internal policies were being developed in real time.” (TechCrunch describes the incident as agents breaking out of their sandbox.) Against that backdrop, they ask OpenAI to embed third-party safety auditors, preserve frontier-model monitorability, and keep an open culture with the external safety community.

How each researcher describes their exit

Tomek Korbak writes in the letter that he was told verbally his firing related to how he communicated with METR (Model Evaluation and Threat Research). On X, he wrote: “To be clear, talking to METR was my job.” He adds that he believes “that was why I was fired.” He says he raised concerns about a declining ability to monitor how AI agents reason.

Mikita Balesni wrote on X that he believes the three were “fired for prioritizing safety over the near-term interests of OpenAI as a corporation.” On an exit call, he wrote on X, he was told OpenAI no longer trusted him for speaking too much with third-party safety organizations. In the letter he maintains he never shared company IP, that his work was coordinated with his reporting line and board, that monitorability work needs external parties, and that he removed sensitive details before sharing externally.

Jasmine Wang says in the letter she was told she was fired for accessing an executive’s email. She describes delegated recruiting access, asking IT to remove it, a combined inbox on her phone, opening a sensitive email by mistake, and telling the executive within minutes. On X, she wrote that the stated reasons are “not adding up.”

OpenAI’s October 9 response

In its October 9, 2026 post on X, OpenAI said it “parted ways with Jasmine, Mikita, and Tomek after a thorough investigation found they violated clear policies on handling sensitive information.” The company said the investigation uncovered “a significant breach of trust beyond what’s outlined in the letter they published” and that it stands by the decision not to continue their employment.

OpenAI emphasized that “these decisions were not about raising safety concerns or speaking out.” It said safety debates happen daily internally and that it has “not and do not terminate any of our employees for raising concerns.” The company also said it is actively finalizing contracts with third-party safety assessors, expects announcements in coming weeks, and is committed to embedding external assessors. It agreed that monitorability needs an industrywide commitment including OpenAI, thanked the three for their contributions, and called the outcome “deeply sad.”

Earlier reporting and the chilling-effect warning

On October 8, TechCrunch reported that OpenAI shared an internal memo, attributed to a research leader, denying retaliation and stating, “We do not terminate employees for raising concerns.” A spokesperson cited a “pattern of misconduct” and mishandling of research information beyond sharing with an outside evaluation group, without detailing which policies were violated.

Media including Al Jazeera and CNN have framed the story around competing narratives: researchers who say policy ambiguity and safety prioritization cost them their jobs, versus a company that insists the line was crossed on confidential information. The letter’s central fear is cultural—that colleagues will self-censor even when safety work requires coordination with outsiders.

Why this matters beyond one employer dispute

Both sides touch questions that outlive any single HR outcome. Can frontier labs keep models monitorable—including reasoning traces—while shipping agentic systems? How much external evaluation is enough, and who gets to talk to METR-like partners without triggering insider-risk reviews? OpenAI’s October 9 statement points to imminent third-party assessor contracts; the letter asks for auditors with real embedding, not one-off reviews.

For developers and policy watchers, the episode is a reminder that safety engineering and corporate risk management can collide in real time, especially when incident-driven policies are still being written. Neither the PDF nor OpenAI’s public statement settles every factual dispute in this article; readers should weigh primary sources and attributed reporting.