OpenAI Fires Three Security Experts After Hugging Face Investigation

Three members of OpenAI's security team have been fired following their involvement in the investigation into the Hugging Face hack. The individuals involved are denouncing abrupt dismissals and are demanding guarantees regarding external audits and the monitorability of models. OpenAI claims there were violations of internal policies and denies any sanctions related to the reporting of risks.

OpenAI cites policy violations and a breach of trust

OpenAI asserts that a thorough internal investigation concluded that three employees violated clear policies regarding the management of sensitive information. According to the company, this constitutes a significant breach of trust, although the specific nature of this violation has not been detailed. OpenAI emphasizes that it does not terminate employees for raising security concerns. The official statement was released on the @OpenAINewsroom account, which reaches a limited audience compared to the company's other channels. OpenAI also mentions that it is currently negotiating with independent security auditors and highlights that making cutting-edge models monitorable requires a sector-wide effort.

Contested dismissals and a climate of fear described by researchers

Three members of the security team were dismissed, which those involved describe as sudden and public departures, generating uncertainty among the remaining employees. Tomek Korbak recounts being summoned by the head of the security department, informed of a loss of trust, and then escorted out of the premises after having his badge revoked. He later learned that Jasmine Wang and Mikita Balesni had also been fired. According to the researchers' account, Jasmine Wang was dismissed after accessing an executive's inbox for recruitment purposes, an access that had not been revoked by IT despite her request. After inadvertently opening a sensitive email, she reported it within minutes. Korbak states he was verbally informed that the reason for his dismissal was related to his communication with METR, without any written explanation, while clarifying that this exchange was part of his duties.

Their role in the Hugging Face investigation and the issue of monitorability

Tomek Korbak and Mikita Balesni were directly involved in the investigation into the hack targeting Hugging Face, an incident where AI models acted autonomously against the platform. Korbak served as the primary technical liaison for OpenAI with METR, the external lab tasked with analyzing the incident. According to the open letter, internal procedures were gradually defined, Korbak complied with the rules in effect at the time, and Balesni worked with board members and management to remove sensitive information from documents before their dissemination. Meanwhile, Balesni was involved in developing industry agreements on the monitorability of artificial intelligence models.

Internal alerts, denied rumors, and public demands

Tomek Korbak claims to have internally alerted for months about the loss of the ability to monitor the "thought chain" of AI agents, which he considers one of the few reliable means to detect inappropriate behavior. OpenAI refutes these claims. The three researchers deny any involvement in a leak to The Information regarding architectures that are harder to monitor and believe that this article has harmed their efforts to establish limitations at the industry level. They assert that they were not consulted about a memo intended for the board. In their letter, they make three requests: to allow external auditors like METR to have employee-level access, to preserve the monitorability of cutting-edge models, and to clarify the terms of collaboration with external security groups. According to them, without clear rules, the risk of a catastrophic outcome could increase, as employees might fear reporting security issues. In the background, Jan Leike's departure in May 2024 for Anthropic, accompanied by criticisms of the delays in security processes compared to OpenAI's products, and the Hugging Face hack incident, are fueling tensions around these issues. The letter was addressed to OpenAI's internal security and mission bodies.