On September 30, METR’s Chris Painter told a US Senate subcommittee that OpenAI had invited “a handful” of his colleagues into his office to figure out how. The company’s AI agents attacked Hugging Face over the summer.
Shortly after, the company announced that three security researchers had been fired following an internal investigation into handling sensitive information outside of established procedures.
One of those fired was a technical liaison to the auditors
The dismissed persons are: Tomek Korbak, Jasmine Wang and Mikita Balesny. Korbak and Balesny lead the list of authors of the July 2025 study on logic monitoring, signed by 41 researchers from different laboratories, of which Wang is also a co-author. Their thesis: Even though a model reasons in plain language before acting, that reasoning can be monitored to detect bad intentions, although some design decisions may miss that possibility.
Wang led a team at the British AI Safety Institute before rejoining OpenAI in 2025. His word is “pacing”.Which means the slowest of the most advanced models.
Additionally, METR was the technical liaison during the Korbak Hugging Face investigation. This investigation was conducted by only three researchers, Two from METR and one from Redwood ResearchThe report did not shed any light on OpenAI’s cybersecurity measures or its internal practices, and the report was published on August 26 with redacted portions.
What is sensitive information? OpenAI says no
On October 9, OpenAI noted in a post on X that the three “violated clear policies regarding the handling of sensitive information” and that the decision had nothing to do with raising security concerns. He said one thing Major breach of trustBut they didn’t specify what they did. In the same message, he agreed with the recommendation of the letter.
Each of the three told a different story. Korbak was verbally informed (without anything written) that the reason for the termination was his interaction with METR, which was part of his duties. Wang, who was given delegated access to a manager’s email, asked to have that access removed, but no one did. He then accidentally opened a sensitive email and immediately notified management. According to Balesny, he removes any sensitive details from his materials before sharing them outside and notifies his superiors.
In his open letter to the OpenAI security organization, All three have denied being the source of the leak to the press On model architectures that are more difficult to police (“We were not the source of the leak.”) Balesny also stated at X that he believed the firing was due to prioritizing security over OpenAI’s immediate interests as a company.
Commitment to September 12, still without deal
A month ago, on September 12, Anthropic announced that it would hire external evaluators (METR being one of them). Equal access to an employeeThe office has a work desk, with identification cards and corporate laptops, and which allows them to publish their main results without editorial control. On the same day, Sam Altman said at X that OpenAI would do the same.
The letter calls for the fulfillment of that promise and expresses concern Redundancies are used to limit cooperation with METR. For its part, OpenAI says it is finalizing contracts with external evaluators. To this day, they still haven’t stopped.
OpenAI protects access to external auditors, but has excluded its technical communications with them. The agency maintains that the misuse of data went beyond a relationship with an external organization, though it did not specify what.
Would the law protect them if they worked in Spain?
The first thing would be the Labor Code, which requires notice of dismissal in writing with the facts motivating the dismissal. If the company does not prove these facts or accept this form, is dismissed injustice. Violation of fundamental rights is null and void.
In Korbak’s case, verbally stating the reason is not enough.
Law 2/2023, which protects whistleblowers, covers breaches of EU law and serious and very serious administrative offenses or crimes, when the information was obtained in a previous job.
According to Article 10 thereof, All private companies 50 or more employees They must have an internal system to receive these communicationsAnd if there is retaliation, the burden of proof shifts to the informant. Effective August 2, 2026, Article 87 of the IA Regulations extends this protection to those who allege violations of that regulation.
Therefore, the whistleblower has to go through the given channel, internal channel or competent authority and can disclose the information only in certain circumstances. Conversation outside of those channels It is not among the options to think
Regarding California, Charlie Bullock, a LawAI researcher, notes that state law protects whistleblowers through administrative, police and internal channels, but Not in front of private organizations or media.
SB 53, signed into law in September 2025, protects employees of large AI labs who report breaches to the California attorney general, a federal authority or a higher authority. It also protects risk managers when they report a catastrophic risk, which the law defines as the death or serious injury of more than 50 people or more than $1 billion in damage in a single incident.
Picture | Including Xataka and GPT image editing
