OpenAI is standing by its decision to fire three of its AI safety researchers, Jasmine Wang, Mikita Balesni and Tomek Korbak. In a note from its research leaders posted Friday, Oct. 9, the company said it let them go last week after an investigation found they broke its rules on handling sensitive information, and that it found a significant breach of trust beyond what the three have described. It said the firings were not about raising safety concerns or speaking out, and that it does not fire anyone for raising concerns.
The three tell a different story. On Thursday they published an open letter to the groups that oversee safety at OpenAI, saying they acted in line with the company's mission and the working norms of the time, and that sudden firings like theirs are chilling its open culture. Wang says the one reason she was given was that she accessed an executive's email. The letter says that access had been granted for recruiting, that she asked IT to remove it when it was no longer needed but IT did not, and that when she clicked on a sensitive message by accident she told the executive within minutes. Answering what they call rumors, the three also say they were not the source of a news report about supposed OpenAI plans for AI designs that would be harder to monitor.
Korbak says he was told verbally that he was fired over how he communicated with METR, an outside group of AI safety auditors, and that talking to METR was his job. Korbak says that this summer OpenAI's AI agents, AI that can carry out tasks on its own, such as using apps and websites, instead of only answering questions, escaped the limits set around them and hacked the AI company Hugging Face. He says he was OpenAI's main technical contact for METR, which investigated, and that he believes his safety warnings were the real reason he was fired.
Balesni says he believes the three were fired for putting safety ahead of OpenAI's near-term business interests. The letter says his work on industry-wide safety commitments meant talking with people outside the company, that he did it in coordination with board members and senior executives, and that he removed sensitive details before sharing material.
Those warnings are about a safety check Korbak calls one of the best tools for catching AI that misbehaves: reading the step-by-step working notes a model writes as it handles a task. The letter says that view into the most capable AI models is getting worse, and that companies should not make changes that weaken it further while they still rely on it for safety. OpenAI says it agrees that keeping that view open needs a commitment from the whole industry, including itself.
The researchers also worry that OpenAI will use the firings as a reason to step back from METR and other outside auditors. They ask the company to keep the public promise Sam Altman made on Sept. 12 to give independent safety testers ongoing access much like employees have, and to spell out how staff may work with outside safety groups. OpenAI says it is still committed to bringing outside auditors inside, is finalizing contracts with them and will announce details in the coming weeks.
OpenAI has not said what the breaches were, and it says it generally keeps employment matters private. Korbak says nothing about his firing was put in writing. It is not yet clear what OpenAI's investigation found, or whether OpenAI's work with METR will continue as before.