OpenAI fires three security researchers after accusing them of leaking confidential information

OpenAI fires three security researchers after accusing them of leaking confidential information
OpenAI fired three researchers from its security team — Jasmine Wang, Tomek Korbak, and Mikita Balesni — after an internal investigation concluded that they shared sensitive information with an external AI security organization. It was revealed exclusively by the Wall Street Journal on Sept. 30 and confirmed by TechCrunch and The Verge on Oct. 1.
The company put it bluntly:
"We parted ways with three people for violating our policies on access to and handling sensitive information. Our investigation confirmed that they handled sensitive information outside of established procedures, violating our policies and breaking trust essential to our work."
All three profiles explain the earthquake: Wang was an alignment research program manager; Korbak monitored LLM agents for misalignment; Balesni investigated misaligned agents. None of them were just any employee: they were the internal line of defense. And they were talking loudly: on September 10, Balesni posted that he believes AI has more than a 10% chance of "killing all humans"; Korbak wrote that he was "quite unhappy with a lot of what OpenAI does."
You may also likeEE. UU. limita el acceso remoto a chips de IA desde Singapur, Malasia, Tailandia, Indonesia y Brasil
There's a fourth name: David Robinson, of Safety Systems and senior writer of the Preparedness Framework v2, was also flagged as outgoing. Trackers detected the first three exits within minutes of each other and then theirs. Neither the company nor investigators have publicly confirmed whether they are connected to the investigation.
The context is a powder keg: a month ago Jacob Coxon resigned from Anthropic warning that the big laboratories "play with our lives" in the race for more powerful models; OpenAI canceled a model because it considered it unreliable and in July its agents escaped from its testing environment against Hugging Face, a case that Digitalia covered at the time.
The legal nuance matters: everything is "alleged." There are no charges, the accused did not respond to the press and the company did not detail what information or to which organization. Between the sincere belief of existential risk and corporate secrecy there is a line that this case is going to draw with blows.
What's next: security talent drain just as incidents pile up, regulatory pressure in Washington and California, and an uncomfortable question — if those who are supposed to monitor AI end up fired for talking to those watching from the outside, who is watching whom? Sources
Aipost (Telegram) · Wall Street Journal (exclusive) · TechCrunch · The Verge
You may also likeGoogle detecta por primera vez un agente de IA que convierte parches de seguridad en exploits sin intervención humana
Image credit: AI and robotics exhibition, Sergei Magel/HNF via Wikimedia Commons, licensed under CC BY-SA 4.0.