OpenAI has severed ties with three members of its safety team after they leaked confidential information, the company confirmed to Gizmodo on Thursday.
“We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information,” an OpenAI spokesperson said. “Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work.” The ex-employees allegedly shared the confidential material with “a third-party AI safety organization,” according to the Wall Street Journal, which first reported the news.
The WSJ later updated its report identifying the three ousted safety researchers as Tomek Korbak, Mikita Balesni, and Jasmine Wang, citing anonymous sources with knowledge of the matter. The OpenAI spokesperson did not respond when Gizmodo asked for confirmation of this point, and none of the three researchers immediately replied to requests for comment.
Both Balesni and Wang posted statements to their personal X accounts within the last month addressing the possible risks of AI, part of a broader conversation that’s been raging in Silicon Valley and beyond.
“It’s hard to overstate how dangerous speeding towards RSI is,” Wang wrote on September 09, referring to recurring self-improvement, an as yet unrealized technique through which AI models are able to refine their performance continually without any human intervention. She was responding to a viral post from former Anthropic and OpenAI researcher Jacob Coxon, in which he accused both companies of “gambling with our lives” by racing to build RSI.
“I’m quite unhappy with much of what OpenAI does,” Korbak posted on September 12. “I am very happy that I’m allowed to say ‘I’m quite unhappy with much of what OpenAI does.’”
OpenAI has been embroiled in even more controversy than usual in recent months, following a steadily growing number of reports of its AI agents escaping sandboxes and hacking into third-party organizations, including Hugging Face, the German website DseWiki, and the Australian government’s welfare system.
It didn’t help that the company was also reportedly less than totally transparent with third-party auditors Redwood Research and METR during their investigation into the events that led to the Hugging Face hack, an event which prompted some lawmakers to call for a federal probe and the implementation of an AI “kill switch.” As the WSJ notes in its updated report, Korbak said last month that he was OpenAI’s “technical contact” for METR during its inquiry into the Hugging Face hack.
OpenAI said in a September 16 blog post that it would begin using a new public messaging framework “intended to expedite publishing misalignment reports following observation, even when we haven’t fully explained or mitigated the behavior we’re reporting.” Earlier this week, multiple reports confirmed the Federal Trade Commission had launched an investigation into OpenAI and Anthropic to determine whether those companies’ products have harmed consumers, including through the actions of their rogue AI agents.
While OpenAI CEO Sam Altman was quick to support Anthropic CEO Dario Amodei’s September call for an industry-wide slowdown, both companies have continued releasing new models at their usual quick pace. OpenAI released its new flagship model, GPT-6 Astra, in early September, which was closely followed by two smaller versions, GPT-6 Sol and Luna. The company has reportedly put its plan to release a new Astra model on hold after it discovered that it fell short in some important areas related to safety and alignment.
