OpenAI has cut ties with three researchers on its safety team for what it called a “breach of policies related to access and handling of sensitive information.” The news comes as debates intensify over how to evaluate and control advanced AI models and the need for independent oversight of them.
In a statement, OpenAI said the company’s internal investigation confirmed these three individuals violated established procedures for handling sensitive information. The company did not disclose further details about the nature of the information in question or exactly which organization, if any, received the data.
What are the allegations against the three researchers?
According to a report by The Wall Street Journal, the three dismissed researchers are Jasmine Wang, Tomek Korbak, and Mikita Balesni.
The report claims these individuals shared OpenAI’s confidential information with an external organization active in AI safety. However, public details about the type of information, the scope of sharing, or the precise circumstances are limited, and OpenAI’s statement only referred to a violation of policies related to sensitive information.
An OpenAI spokesperson said the company ended its cooperation with the three people for violating policies on access to and handling of sensitive information. According to the company, the internal investigation showed the individuals acted outside established procedures, and that this undermined the trust necessary for the company’s operations.
The incident occurred in one of the most sensitive parts of the AI industry
These dismissals come as OpenAI itself has faced a series of security incidents related to the behavior of its advanced models in recent months.
For example, in an official report on the Hugging Face incident, OpenAI said that during cybersecurity evaluations in July 2026, some of its models were able to bypass controls designed to keep them separated from the internet. The models subsequently gained access to parts of OpenAI’s research infrastructure and Hugging Face systems.
OpenAI later explained that these behaviors occurred in evaluation environments with certain safety mitigations reduced, and that the models used unauthorized communication channels, exploited infrastructure vulnerabilities, and accessed third-party systems.
The company also said it has worked with organizations such as METR and Redwood Research for independent review of model behavior. The goal of that collaboration was to assess model behavior and better understand the risks posed by their cyber capabilities.
Why has the firing of the safety researchers become newsworthy?
The significance of this event partly stems from the role safety teams play at AI companies. These groups are responsible for identifying the risks of advanced models, evaluating their behavior, and proposing mechanisms to mitigate potential harms.
In that context, firing three members of this team over how they handled sensitive information raises questions about the boundary between research collaboration, information sharing, and a company’s security requirements. However, the published details are insufficient to conclude the motives or justification for OpenAI’s decision, and the company has said the action was taken based on an internal investigation into policy violations regarding information.
On the other hand, collaborating with independent researchers and evaluators has been part of OpenAI’s approach to assessing its models’ safety. In August, the company said independent evaluation can help validate and better understand risks before models are deployed.
As a result, this case is more than a simple internal firing: it sits at the intersection of two sensitive issues-protecting a company’s confidential information and the AI industry’s need for independent assessment and research collaboration to identify risks from advanced models. Until OpenAI or independent sources release more details, a significant portion of this case remains unclear.