OpenAI parts ways with three researchers over sensitive-information handling

Why it matters

OpenAI is expanding external safety assessments while policing access to confidential research. What the researchers shared, and with whom, will clarify how the company draws that boundary.

OpenAI has parted ways with three researchers after an internal investigation found they mishandled sensitive company information, the company said on October 1st. The Wall Street Journal reported that the researchers allegedly shared confidential information with a third-party AI safety organization. OpenAI has not named the researchers, the organization, or the material involved. Max Zeff reported the departures; The Wall Street Journal's report described the alleged sharing.

An OpenAI spokesperson told CBS News that the company's investigation found the employees handled information outside established procedures and that the probe uncovered a pattern of misconduct among people with access to confidential research. The company described their departures as a response to policy violations. It did not say whether they were fired, resigned, or left under another arrangement, and did not specify what information was involved.

The account arrives amid a closely related policy push. On September 22nd, OpenAI published principles for third-party safety assessments, saying independent evaluators need meaningful access to training, evaluation, and deployment information to test the company's safety claims. OpenAI said those assessments can involve access to technical safeguards, internal deployment data, and model reasoning. It also said labs and assessors share responsibility for security, and that evaluations require strong safeguards for sensitive information.

That is the operational tension in this personnel case: outside scrutiny depends on access to information that OpenAI treats as confidential. OpenAI's statement says the researchers crossed its internal rules; the public account does not establish what they shared or whether the recipient was acting as an evaluator, an independent researcher, or in another capacity. Those details would determine how the incident fits the company's stated approach to external review.

OpenAI's own disclosure framework, published on September 16th, says the company intends to report qualifying examples of model misalignment, including cases involving third parties, while allowing security and legal responsibilities to shape the timing and detail of disclosure. The framework sorts cases into tracks based on the amount of investigation needed and says complex cases involving third parties may take longer. It also says employees can escalate disagreements about whether an incident should be disclosed to senior leadership. That framework covers model behavior, not employee conduct, but it illustrates how OpenAI is trying to balance public scrutiny with sensitive information and third-party obligations.

The timing is significant. OpenAI's September 16th post included six examples of concerning model behavior from training or evaluation. On September 29th, the company also delayed the release of a model after safety concerns, according to The Associated Press. The employee departures are a separate matter, and the available reporting does not connect them to either development. Taken together, however, the events put safety research and the handling of internal evidence under increased scrutiny.

OpenAI's founding statement, published in 2015 by its founding group, including Sam Altman, described a goal of advancing AI for broad human benefit and said the organization expected to collaborate with other institutions. Its present-day work with external assessors makes such collaboration a practical requirement, while its latest statement emphasizes that access to internal research depends on trust and policy compliance. The company has yet to describe where it draws the line between permitted evaluation and prohibited disclosure in this specific case.

The distinction will matter to outside researchers deciding what access they can receive, and to OpenAI employees handling material used to evaluate powerful models. For now, the confirmed account is narrow: three researchers have left, OpenAI says its investigation found mishandling of sensitive information, and the WSJ reported an alleged disclosure to an unnamed AI safety organization. The nature of the information and the recipient's role remain central to understanding the incident.