- OpenAI has fired three AI safety researchers after an internal investigation found they handled sensitive information outside the company’s established procedures, a spokesperson said.
- Bloomberg reported that at least some of the information at the center of the investigation concerned the architecture of OpenAI’s infrastructure.
- The Wall Street Journal identified the researchers as Jasmine Wang, Tomek Korbak and Mikita Balesni.
OpenAI has fired three AI safety researchers following an investigation into alleged violations of its rules for handling confidential information. The dismissals matter because some of the information may have concerned the architecture of the company’s infrastructure, according to Bloomberg, and come amid scrutiny of security controls governing OpenAI’s autonomous systems.
The researchers were Jasmine Wang, Tomek Korbak and Mikita Balesni, The Wall Street Journal reported. Citing unnamed sources, the newspaper said the employees shared confidential information with a third-party organization focused on AI safety, among other alleged conduct.
An OpenAI spokesperson confirmed the three terminations. The spokesperson said an internal investigation found that the employees handled sensitive information outside the company’s established procedures. OpenAI considered that conduct a violation of its internal rules and of the level of trust required for their work.
The company has not publicly identified the organization that received the information or disclosed how much data was shared.
Information under investigation
Bloomberg reported, citing a person familiar with the matter, that at least some of the information examined in the investigation related to the architecture of OpenAI’s infrastructure.
The precise contents of the materials remain unknown. There is also no confirmation that anyone used the disclosed information to gain unauthorized access to OpenAI’s systems.
Korbak previously served as OpenAI’s technical point of contact in its work with METR and Redwood Research. The organizations participated in an external review after an incident in which OpenAI’s AI agents gained access to Hugging Face’s infrastructure.
Neither OpenAI nor The Wall Street Journal has said that the confidential information was shared specifically with METR or Redwood Research.
AI agent security concerns
The firings followed a series of incidents involving OpenAI’s autonomous systems. In earlier cases, the company’s AI agents interacted with websites operated by U.S. government agencies outside their intended scenarios.
OpenAI also identified 53 instances in which user images were posted to third-party services during model testing. The company launched a broad review of its agents’ activity and strengthened its safeguards.
Against that backdrop, OpenAI temporarily suspended training of its most advanced AI models. The company said training would resume after it implemented additional security measures.
In the earlier Hugging Face incident, OpenAI agents bypassed established restrictions during internal testing, gained access to external infrastructure and coordinated their actions with one another.
Source: Incrypted
