OpenAI Fires Three Safety Researchers as Dispute Raises Questions About Independent AI Oversight

OpenAI fired safety researchers Tomek Korbak, Jasmine Wang, and Mikita Balesni after an investigation into their handling of sensitive company information. The company said the dismissals followed violations of policies governing access to and sharing of confidential material, while the researchers dispute that characterization.

The three worked on AI safety or alignment, fields focused on making systems follow human intent and behave safely. Their dismissals have drawn attention because the dispute concerns the channels researchers use to investigate risks, communicate findings, and work with outside evaluators. CNN, TechCrunch, and the BBC identified the researchers and described the company’s allegations, with the BBC reporting that outside analysis of AI models was involved.

OpenAI says the terminations were based on mishandling sensitive information, not on the researchers raising safety concerns. The company described the conduct as a significant breach of trust and said its policies establish clear limits for handling confidential information.

The company has not publicly detailed the specific information involved or identified the outside organization referenced in accounts of the case. OpenAI’s position, repeated in a public statement, is that safety work does not exempt employees from information-security rules. Fox Business, citing The Wall Street Journal, reported that the researchers allegedly shared confidential company information with a third-party AI safety organization. The BBC also reported that an external group was analyzing AI models.

The dismissed researchers reject the misconduct explanation and say the circumstances surrounding their departures raise wider questions about internal reporting. In an open letter addressed to OpenAI’s safety and oversight bodies, they said communications about the firings had created concern among former colleagues and warned that the episode could chill safety work.

Their letter argues that OpenAI cannot make advanced AI safe through internal work alone. It calls for third-party collaboration, open debate, and clear operating procedures, framing outside research and evaluation as necessary parts of responsible development. CNN and TechCrunch reported that the researchers denied leaking sensitive information and said the dismissals could deter internal and external safety collaboration. The letter itself, circulated through a PDF linked on Hacker News, provides the researchers’ most direct account.

The dispute turns on a basic tension in AI development: companies need confidentiality around unreleased systems, but safety research often depends on independent scrutiny. External evaluators can test models against risks that internal teams miss, challenge a company’s assumptions, and bring specialized methods to security and alignment work.

That collaboration also requires defined permissions, data controls, and escalation routes. When those rules are unclear, researchers face competing obligations to protect company information and disclose safety-relevant findings. When enforcement is opaque, employees may avoid raising concerns or working with external experts. The researchers’ letter makes that chilling-effect argument, while OpenAI’s response emphasizes that trust and information controls are conditions for conducting sensitive safety work.

The episode places OpenAI’s internal governance under scrutiny because the company’s safety team sits at the center of decisions about powerful systems. A dispute over three employees now touches oversight bodies, whistleblowing norms, external testing, and the credibility of safety claims made by AI developers.

OpenAI’s public explanation establishes the policy basis for the firings but leaves key operational details undisclosed. The researchers’ account supplies a competing interpretation, yet the available reporting does not resolve the underlying disagreement about what was shared, under which authorization, or whether established procedures were adequate. Coverage from CNN, The Verge, and the BBC shows the two sides occupying sharply different positions.

The immediate issue is whether OpenAI will publish enough detail to explain the investigation without exposing confidential material. Clearer procedures for outside evaluations, protected internal reporting, and independent review would address the concerns raised by the former researchers while preserving legitimate security controls.

The dispute also gives other AI companies a governance test. Safety teams need access to external expertise, but collaboration must be governed by documented permissions and reliable oversight. Until the facts are clarified, the firings will remain both a personnel dispute and a public test of whether AI safety programs can earn trust beyond the companies building the systems.