Could OpenAI firings chill outside collaboration on AI safety work?

All three worked on teams devoted to AI safety and to making OpenAI’s models follow human intentions. Two of them were closely involved in the investigation of the Hugging Face hack, in which a swarm of OpenAI agents escaped their test environment and breached outside systems.

• Tomek Korbak was the technical point of contact for outside investigators from two safety groups, METR and Redwood Research, who were allowed to examine internal records from the incident.
• Mikita Balesni worked on monitorability, the ability to watch an AI model’s reasoning, which he says required lots of contact with outside experts. The letter says he had backing from board members and executives.
• Jasmine Wang coined the term “pacing,” the idea of slowing the most advanced AI work so safety research can catch up. More than a thousand AI staff later invoked it in a July open letter calling for a slowdown.

About a week after the firings, the three published an open letter to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. Their central point was that behavior that counted as normal a month earlier had suddenly become a firing offense, leaving former colleagues afraid to speak up or work with outsiders.

They argued that safety researchers see risks first and depend on outside experts to deal with them, so the freedom to collaborate without fear is itself a safety tool. They wrote that OpenAI is “not a normal company.”

The letter also denied two things. The three said they had nothing to do with a leak to The Information about OpenAI’s newest model, and that they never dealt with outside parties beyond what their jobs required. They asked OpenAI to keep its promise to place third-party auditors inside the company, to protect the ability to monitor its most advanced models, and to keep an open culture with the wider safety community.

OpenAI fired me last week, along with two of my safety colleagues. I was given one reason: that I accessed an executive's email. I want to say this plainly, because too many of OpenAI’s history is smoke and mirrors when people disappear: https://t.co/RC4fm2nV95

— Jasmine Wang (@j_asminewang) October 8, 2026

OpenAI says the firings had nothing to do with safety concerns. An internal memo from a research leader praised the trio’s work and said the company does not fire people for raising concerns. A spokesperson went further, saying an investigation found a pattern of misconduct in clear breach of its rules on research information, going beyond any sharing with an outside evaluation group, and that the violations happened more than once.

The company has not said exactly which policies were broken. It did say publicly that it agrees with the letter’s recommendations and remains committed to bringing in outside evaluators, a promise CEO Sam Altman made on Sept. 12 when he said OpenAI would follow Anthropic in widening access for third-party testers.

Later reporting filled in what the original letter left out. In late October, The Wall Street Journal reported that OpenAI’s security team had asked the three to open their personal phones during a leak investigation. The inquiry followed a Sept. 1 story in The Information claiming Astra, OpenAI’s latest model, used a new technique that made it stronger but showed less of its reasoning.

On the phones, investigators found that some of the researchers had turned on disappearing messages in Signal chats with Ryan Greenblatt, a former chief scientist at Redwood Research who now works at METR. OpenAI reportedly suspected him of the leak. A week later, the three were fired.

Each researcher tells a slightly different story:

• Korbak says he was told there had been a breach of trust because he dealt with a third party through unapproved channels. He called disappearing messages a basic security habit in tech.
• Balesni says he was told OpenAI no longer trusted him because he spoke too much to outside safety groups, and he insists he never shared company secrets.
• Wang says she was fired for opening an executive’s email. She says that access was given to her for recruiting work, she asked IT to remove it, and when she opened a sensitive message by mistake she told the executive within minutes.

An OpenAI spokeswoman said Signal use was not approved but was not the only reason. Greenblatt denies leaking anything and wrote on X that OpenAI had made false accusations about him. He says he saw only a cleaned-up version of an internal memo that had been approved for sharing with him.

For ordinary users, this fight is about who gets to check the work of companies building the most powerful AI. Independent groups like METR found out how far OpenAI’s agents went in the Hugging Face hack only because they were let inside. Balesni has said he fears OpenAI could use the firings as a reason to cut ties with METR.

Both sides now say they want outside evaluators. The open question is whether OpenAI’s staff believe they can work with those evaluators without risking their jobs. Wang put the warning bluntly: anyone who raises concerns could be next, without being told why.

Balesni and Korbak gave their first interview in late October, describing life inside OpenAI during the Hugging Face investigation and arguing that the company’s rebuttal doesn’t add up. The next test is concrete: whether OpenAI signs the outside evaluators it has promised and how much access they get.