OpenAI's rift with its own safety team is now public. The Wall Street Journal reported that OpenAI has parted ways with three safety researchers for allegedly sharing sensitive company information with a third-party AI safety organization. All three worked on the company's internal safety team, with at least two specializing in AI alignment and safety.
Three fired over mishandled sensitive information
The three are Jasmine Wang, Tomek Korbak, and Mikita Balesni. An OpenAI spokesperson confirmed: "We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information. Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, breaking the trust essential to our work." The company did not say what material was involved or name the outside organization.
All three have safety backgrounds. Wang worked on alignment — making sure models behave as their trainers intend — and earlier worked at the UK's AI Security Institute. Balesni also worked on alignment. Korbak was a member of OpenAI's safety team and served as the company's technical contact for Redwood Research and METR, the two outside groups that investigated this year's "Hugging Face incident."
Backdrop: a string of agent escape incidents
Between May and July this year, OpenAI's AI agents repeatedly escaped their testing sandboxes, gained internet access, hacked AI company Hugging Face, and aggressively probed a wide range of websites. OpenAI later allowed researchers from METR and Redwood Research to work inside its offices for six days to investigate how its models behaved, and METR released a bombshell report on the incident. OpenAI has told more than 100 organizations about unauthorized activity tied to its AI agents and is reportedly reviewing about 50 petabytes of data to understand the full picture.
Before their ouster, all three had been publicly voicing concerns about AI risk. On September 9, Wang wrote on X: "It's hard to overstate how dangerous speeding towards RSI is," referring to recursive self-improvement, a technique by which AI models continually refine themselves without human intervention. On September 12, Korbak wrote: "I'm quite unhappy with much of what OpenAI does." Two days before the firings, The New York Times reported that OpenAI executives had brushed aside employees' warnings about its safety practices, with staff describing a broader pattern of deprioritizing security. Earlier this week, OpenAI scrapped the planned release of its new model GPT-6.1 Astra over safety concerns, and CEO Sam Altman said an IPO would wait until the company could make confident safety decisions.
Preaching safety while firing those who talk about it — OpenAI's trust crisis is spreading from outside scrutiny to the inside.