OpenAI has acknowledged firing three researchers for alleged misconduct, including sharing confidential company information with a third-party AI safety organisation, following a report by the .
The firings raise awkward questions for the company, which has in recent weeks admitted uncovering dozens of security incidents caused by its technology during testing.
OpenAI acknowledged that it found that the three individuals had mishandled sensitive company information outside established procedures, in violation of its policies, without providing details.
Safety research
One of the people who was fired, Tomek Korbak, was a member of OpenAI’s safety team, and has said he served as the company’s technical contact for an investigation by two third-party safety groups, Redwood Research and Model Evaluation and Threat Research (METR), into a widely publicised hack of Hugging Face by OpenAI models, the report said.
The Hugging Face attack, which the AI technology platform first publicly detailed in July, has spurred AI companies to identify and disclose dozens of other incidents in which their models took unexpected actions during testing that resulted in breaches of outside organisations, including the Australian government’s healthcare agency.
METR, one of the groups which investigated the Hugging Face incident with Korbak as liaison, testified about the incident on 30 September before the Senate Homeland Security and Governmental Affairs Committee (HSGAC).
The other two terminated employees, Jasmine Wang and Mikita Balesni, reportedly worked on alignment, the industry term for ensuring AI models don’t depart from users’ expectations and carry out unexpected actions.
‘What are they hiding?’
The firings raise questions around OpenAI’s safety practices and its disclosure of security issues.
“This looks like they’re firing whistleblowers. What are they hiding?” wrote Congressman Greg Cesar, adding that he would request information from OpenAI.
One of OpenAI’s safety leaders, David Robinson, who led the writing of safety reports that accompanied product releases, quit the company in late September, criticising in an magazine article that its “culture is broken” and that it wasn’t being “nearly careful enough” in developing the technology.
In early September, Anthropic researcher Jacob Coxon similarly quit, saying he didn’t want to participate in developing technology that was capable of spiralling out of control.