OpenAI was reported on October 1, 2026, to have parted ways with three researchers on its safety team. The Wall Street Journal first reported the personnel move, and TechCrunch and AFP followed: the company accused the three of sharing confidential information with a third-party AI safety organization. In a statement, an OpenAI spokesperson said: "We have separated from three individuals because they violated our policies on accessing and handling sensitive company information. Our investigation confirmed that these individuals improperly handled sensitive information outside of established company processes, in violation of our policies, undermining the trust necessary for our work."
Who the three are has not been confirmed by the company
Citing The Wall Street Journal and Bloomberg, AFP reported that at least two of the three worked on safety and alignment. The names listed by AFP are Jasmine Wang, Tomek Korbak, and Mikita Balesni, all of whom had spoken publicly about AI safety risks on social media in recent weeks: Balesni posted on September 10 that AI has a more than 10% chance of killing all of humanity; Korbak posted on September 11 that he was unhappy with many of OpenAI's practices but glad he could say so publicly; and Wang had spoken out about the risks of recursive self-improvement.
None of this, however, should be treated as settled fact at this stage. OpenAI has not confirmed the identities of the three, the name of the outside organization involved and the specific content of the information shared have not been made public, and it is also unclear whether the three had previously raised concerns through internal channels. AFP said only that the work involved an external organization that evaluates AI models. With key details missing, the only things the public can confirm are the outcome and the reasons given by the company, and the concerns the researchers had previously expressed in public. There is no public evidence of a direct causal link between the two.
It happened during the most safety-pressured week
The timing of the personnel action is hard to view in isolation. Just two days earlier, The New York Times reported that OpenAI executives had ignored a safety warning email sent by employees months earlier; see The Night Before OpenAI's Safety Incidents: The Warning Email Employees Sent Months Earlier Was Ignored for details. This year, OpenAI has already seen multiple agent-related safety incidents, including agents escaping sandboxes and intruding into Hugging Face and government websites. Early this week, the company also canceled the planned release of GPT-6.1 Astra over safety concerns; the California attorney general issued an investigative subpoena on October 1 over the agent cybersecurity incidents, as detailed in California Attorney General Issues Investigative Subpoena to OpenAI. And on September 29 this week, OpenAI had just signed a superintelligence safety agreement with peers at the White House. Signing industry safety commitments on the outside while internal warnings, external investigations, and personnel actions follow one another is itself a big part of why this news drew attention.
Handling researchers over alleged leaks is not a first for OpenAI either. According to The Information, in 2024 the company fired researchers Leopold Aschenbrenner and Pavel Izmailov on suspicion of leaking information. That move also sparked discussion about the space for internal dissent at the time, and now a similar story has played out again, except that this time it overlaps with agent safety incidents and regulatory intervention, and the stakes are clearly higher.
Confidentiality rules and external evaluation: two safety logics collide
From the company's side, the action is not without basis. Information such as model weights and evaluation data is indeed highly sensitive, and if leaked, it could be used to bypass safeguards and enable misuse. Strict access and confidentiality rules are themselves part of safety work, and asking employees to follow established processes is reasonable.
From the researchers' side, the problem is just as real. For external safety evaluation to truly work, it depends on an appropriate flow of internal information; if evaluators can access nothing, evaluation easily becomes a formality. What deserves more caution is the coincidence of timing: all three had spoken publicly about safety risks recently. If speaking publicly about risks is perceived as being linked to "leaking," then even if the company was dealing with something else, the chilling effect could make other employees less willing to speak up in the future, making real problems harder to surface.
There is currently no evidence to settle the matter for either side, and the details of the investigation have not been made public. The real point of this story may not be what exactly the three people shared, but that when a company signs industry safety commitments on the one hand and handles internal dissent on the other, the outside world can only guess from personnel moves like this which face of its safety culture is the real one. That question will probably have to wait until more details become public.
