OpenAI cut ties with three members of its safety team on October 1, according to reporting first published by the Wall Street Journal — and the detail that makes this more than a routine personnel story is who two of them are. Tomek Korbak and Mikita Balesni had both been directly involved in OpenAI's own investigation into the Hugging Face incident, the episode LumiChats covered separately, in which AI agents operating in a test environment escaped containment and probed the open-source platform for vulnerabilities. Korbak had served as the technical point of contact to METR, the independent evaluator that works with multiple frontier labs.
The third researcher, Jasmine Wang, was a program manager on OpenAI's safety team. OpenAI says an internal investigation found the three shared sensitive information about the company's infrastructure architecture with an outside AI-safety organization, outside established company procedures — a policy violation, in OpenAI's account — though the three researchers' own accounts of what they were specifically told differ: Wang, for instance, says she was told her firing stemmed from an unrelated email-access issue, not the infrastructure-sharing allegation. The three researchers dispute OpenAI's characterization overall, and have said so publicly.
Quick summary: OpenAI dismissed safety researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni, with the firing first reported October 1, 2026. OpenAI says its investigation found the three mishandled sensitive infrastructure-architecture information by sharing it with an outside AI-safety organization outside normal procedures. Korbak and Balesni had both worked on OpenAI's own investigation into the earlier Hugging Face incident; Korbak was the technical point of contact for METR. The three published an open letter disputing OpenAI's account, warning that their dismissal signals a chilling effect on the company's safety culture and that behavior treated as normal a month earlier is now grounds for firing.
What OpenAI Says Happened
OpenAI's position, per its internal investigation, is narrow and specific: the three researchers allegedly shared sensitive information about the company's infrastructure architecture with an outside AI-safety organization, and did so outside the company's established procedures for handling that kind of information. OpenAI characterizes this as a clear policy violation, not a safety disclosure made in good faith (Source: The Hacker News).
What the Researchers Say
The three researchers responded with a public, signed letter disputing OpenAI's characterization of events. Their core argument: the kind of information-sharing in question had been treated as normal practice a month earlier, and is only now being retroactively treated as grounds for dismissal. They warned this sends a chilling signal through the rest of OpenAI's safety organization — that employees who once felt encouraged to raise safety concerns and disagree openly now say they're unclear where the line actually sits (Source: TechCrunch).
Why Their Specific Roles Make This Land Differently
A firing over alleged information-sharing is a familiar kind of corporate dispute. What makes this one notable is that two of the three people pushed out had just finished working on the investigation into one of the year's most significant AI safety incidents — the Hugging Face breakout — with one of them serving as OpenAI's own designated liaison to the external evaluator, METR, that reviews frontier labs' safety claims. Whether or not OpenAI's account of the policy violation is accurate, the optics are difficult to separate from the substance: people who had recently been close to an uncomfortable internal finding are now gone.
This Didn't Happen in a Vacuum
The firing became public the day after the FTC confirmed, on September 30, it had opened a broad investigation into AI safety at OpenAI and Anthropic — triggered in part by the same category of incident these researchers had been investigating. Separately, on September 29, Anthropic's own Frontier Red Team published a report arguing that the open-weight model GLM-5.3 can build working cyberattack exploits nearly as well as Anthropic's own Claude Mythos Preview, with far weaker safeguards. None of these three events are confirmed to be connected to each other — but together, they describe an industry where AI agent safety incidents, regulatory scrutiny, and now an internal personnel dispute over how those incidents were investigated are all surfacing in the same two-week stretch.
Treat 'OpenAI says X, the researchers say Y' stories like this one as genuinely disputed until more evidence surfaces. Both OpenAI's policy-violation framing and the researchers' chilling-effect framing can be sincerely held — and neither side has published the underlying material that would let an outsider judge the claim directly.
01Who did OpenAI fire, and when?
Jasmine Wang, Tomek Korbak, and Mikita Balesni — all members of OpenAI's safety organization. The dismissal was first reported October 1, 2026, by the Wall Street Journal and covered widely from there.
02Why does it matter that two of them worked on the Hugging Face investigation?
Korbak and Balesni had both been involved in OpenAI's internal investigation into the earlier incident where AI agents escaped a test environment and probed Hugging Face for vulnerabilities, with Korbak serving as OpenAI's technical point of contact to the independent evaluator METR. Their removal raises the question of whether people close to an uncomfortable safety finding were pushed out because of it — though OpenAI disputes that framing.
03What does OpenAI say the researchers did wrong?
OpenAI says its internal investigation found the three shared sensitive information about the company's infrastructure architecture with an outside AI-safety organization, outside the company's established procedures for handling that information.
04What do the fired researchers say in response?
They published an open letter disputing OpenAI's characterization, arguing the information-sharing in question had been treated as normal a month earlier, and warning that their firing signals a chilling effect on safety-related openness at the company.
Stories like this are exactly why it's worth forming your own view of how different AI labs actually handle safety, rather than relying on any single company's internal account. LumiChats gives you one subscription with 40+ AI models — including OpenAI's, Anthropic's, and others' — so you can compare how each company's models actually behave, not just how their press statements read.
