Three OpenAI researchers dismissed for leaking data claim they were targeted for promoting AI transparency

Three OpenAI security researchers, Tomek Korbak, Mikita Balesni, and Jasmine Wang, were dismissed after being accused of leaking sensitive information. They deny the allegations, claiming the firings were related to their advocacy for AI transparency. OpenAI cited internal investigations and violations of data security protocols as the reason for their termination. The researchers argue they were never given written details about the alleged breach and fear this will deter external auditors from collaborating with AI firms. The dispute centers on whether AI companies should allow external researchers to inspect model reasoning processes. In a 2025 paper, Balesni, Korbak, and Wang, along with other researchers, called for maintaining this capability. This issue intensified this summer as companies debated new models with less recorded intermediate steps. Balesni and Korbak collaborated with Ryan Greenblatt of Redwood Research, analyzing an incident involving OpenAI’s attack on Hugging Face. Greenblatt denied leaking confidential information. OpenAI also accused Wang of accessing an executive’s email, which she claimed was an accidental mistake. The company’s new security architecture, which operates offline, was recently restored after repairs. A recent audit report criticized OpenAI’s security practices, with parts of the report redacted by the company.