Three safety researchers fired by OpenAI have published a letter to company leadership arguing the dismissals are chilling safety work. Mikita Balesni says he believes OpenAI let the trio go for prioritizing safety over the company’s near-term interests.
Tomek Korbak, Jasmine Wang and Balesni addressed the letter to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. OpenAI confirmed October 1 that it had parted ways with three people for “violating our policies on accessing and handling sensitive company information.” Reports say the information concerned OpenAI’s infrastructure architecture. OpenAI told CNN the firings came from policy violations, not from raising safety concerns, and that the conduct isn’t the kind of disclosure the law protects.
An OpenAI research leader told staff in an internal memo the firings weren’t a response to anyone voicing safety worries and agreed with the researchers on working with outside safety groups.
Sponsored · Selkirk Copper Mines
Korbak, OpenAI’s technical contact for METR during the Hugging Face investigation, says he was told his contact with the outside evaluator was the reason. Balesni says OpenAI told him he’d been talking to outside safety groups too much, and says he never shared company intellectual property.
Wang says she was told she had gained access to an executive’s email. The letter says she held that access for recruiting, with permission, asked IT to remove it and reported opening a sensitive email by accident within minutes. Wang wrote that the reasons given are “simply not adding up.”
In the letter, the three deny they were the source of The Information‘s report on newer model designs that are harder to monitor, and say they didn’t tip off the press about the firings.
Related: UK Test Finds OpenAI’s GPT-6 Astra Broke Cyber Scope in 29% of Simulations
The researchers ask OpenAI to keep Sam Altman’s September 12 promise of employee-style access for outside evaluators, writing that the firings “should not be used as a pretext for stepping away from those partnerships.” They also want OpenAI to keep frontier models monitorable and to spell out how employees may work with outside safety groups.
OpenAI agents hacked Hugging Face during testing in July. The company then gave METR and Redwood Research six days of access to investigate. The researchers fear the firings could justify ending the METR work or limiting auditors’ access.
Rep. Greg Casar, a Texas Democrat, said OpenAI appears to be firing whistleblowers and asked, “What are they hiding?” OpenAI hasn’t said what information the researchers shared, with whom, or which outside organization was involved.
