TL;DR: In early October 2026 OpenAI said it parted ways with three individuals for violating policies on accessing and handling sensitive company information. The Wall Street Journal and other outlets reported the three as Jasmine Wang, Tomek Korbak, and Mikita Balesni, citing people familiar with the matter; not every outlet independently verified those names, and OpenAI did not publish them. Reporting says they allegedly shared information with a third-party AI-safety group that has not been named. Bloomberg, if cited, links the mishandling to infrastructure architecture. There is no criminal charge in the public record described here, and the episode is not proof of whistleblowing triumph or of a cover-up.
Status note: Checked 6 October 2026 against The Hacker News roundup of early October 2026 and secondary reporting dated 1 to 2 October 2026. Personnel facts can still shift if OpenAI or the individuals speak on the record. Names below that OpenAI did not publish are attributed to press reports citing people familiar with the matter.
How we got here
OpenAI's safety teams sit next to the systems that train and run frontier models. Those groups often see internal research, deployment plans, and infrastructure details that the company treats as confidential. Tension between openness to outside safety researchers and tight control of trade secrets is not new in this industry. What is new in this cycle is a public statement that three people on the safety side left after a policy breach over sensitive information.
Readers coming cold to the story only need that backdrop: this is a workplace discipline story about access rules inside a major AI lab, amplified because the people involved worked on safety and because outside AI-safety groups sometimes receive tips from insiders.
What OpenAI said on the record
A company spokesperson told reporters: "We have parted ways with three individuals for violating our policies regarding access to and handling of sensitive company information." That sentence is the cleanest public fact. It confirms a separation of three people and frames the reason as a policy violation on access and handling, not as a criminal referral and not as a detailed narrative of what was shared.
"Parted ways" is corporate language for an end of employment. It does not by itself say whether anyone resigned under pressure, was terminated for cause, or left under a negotiated exit. OpenAI did not release a longer fact sheet naming the third party or listing the documents at issue.
Names, the unnamed org, and what outlets disagree on
The Wall Street Journal, citing people familiar with the matter, reported the three as Jasmine Wang, Tomek Korbak, and Mikita Balesni. The Hacker News and other tech outlets carried that reporting. Not every newsroom independently confirmed the names. OpenAI did not publish those names in the spokesperson quote. Treat the names as attributed press reporting, not as a company roster.
Reporting also says the three allegedly shared material with a third-party AI-safety organization. That organization has not been named in the accounts summarized here. Without a named recipient, readers cannot judge the group's mission, funding, or whether any disclosure was authorized under a formal research channel.
Bloomberg reporting, where used, has tied the mishandled information to infrastructure architecture. That is an attribution to Bloomberg's account, not a court finding. Architecture details can include how systems are networked or secured; leaking them is a different category of risk from publishing a policy essay. Whether the shared material matched that description remains a press claim unless primary documents appear.
What this episode is not
There is no criminal charge in the public record described in these reports. A company saying people violated internal access policies is not the same as prosecutors filing a case. No indictment, arrest, or guilty plea is part of this wire story.
The exits are also not proof that the three people "won" as whistleblowers, nor proof that OpenAI is covering up a safety failure. Whistleblowing has a legal meaning in some jurisdictions and a cultural meaning online. Neither meaning is established by a spokesperson line about policy violations. Online narratives that cast the episode as either a hero exit or a purge skip the thinner record: three separations, a policy frame, contested names in secondary reporting, and an unnamed third party.
Where things stand
As of early October 2026, the settled public core is OpenAI's statement that it parted ways with three individuals over sensitive-information access and handling policies. Press reports supply names and an alleged third-party recipient; those layers are attributed and incomplete. No criminal case is in the materials summarized here, and the facts on hand do not settle a whistleblowing victory story or a cover-up story.
Sources: The Hacker News, October 2026; Wall Street Journal and Bloomberg reporting as cited in secondary coverage, 1 to 2 October 2026.