OpenAI Defends Firing of Three Safety Researchers Over Breach of Trust

OpenAI defended its decision to fire safety researchers Jasmine Wang, Tomek Korbak and Mikita Balesni, saying an internal investigation found a significant breach of trust. The statement, reported by The Verge, names the three and rejects the idea that the dismissals were retaliation.
OpenAI said the researchers violated clear policies on handling sensitive information. It said the dismissals were not because the researchers spoke out about the company and AI safety concerns. The details stop there.
According to the same report, OpenAI said its investigation uncovered breaches beyond what was outlined in the researchers' own letter. It provided no details. That distinction matters.
The researchers published an open letter on Thursday urging OpenAI to be more transparent about the decision. In it, they said they believed they were fired for raising safety concerns. They also said they had acted in line with OpenAI's mission and working norms of the time.
Earlier reporting had framed the dismissals around mishandling of confidential material. The Wall Street Journal reported on Oct. 1 that OpenAI had parted ways with three researchers for alleged misconduct including sharing confidential company information with a third-party AI-safety entity. TechCrunch separately reported that the company cut ties after an internal investigation found mishandling of sensitive company information.
CNN reported on Oct. 8 that the three said they were fired and openly questioned the reasons given for their dismissals. The report published Oct. 9 from OpenAI is a direct response to that public challenge.
The broader context here is familiar to anyone who has operated access controls around frontier research. Safety work often requires broad visibility into models, evaluations and incident data. Production deployment requires narrow distribution. Those two imperatives collide.
In my view, the core dispute is less about safety as a principle and more about process for handling safety-relevant information. OpenAI is asserting a policy boundary was crossed. The researchers are asserting their actions fit the mission and norms as they understood them. Both claims can feel internally consistent to the people making them when disclosure rules are ambiguous or evolving.
Looking at what this means for technical teams, the lesson is procedural. Clear classification, explicit approval paths for external sharing, and written pre-publication review reduce ambiguity. Without them, even well-intentioned coordination with external safety groups becomes a trust problem. With them, disagreements about risk become reviewable.
Worth flagging: opacity cuts both ways. A company that cites breaches beyond a public letter without describing them protects an investigation but leaves external researchers unable to calibrate behavior. Researchers who appeal to working norms of the time highlight how fast those norms shift as capabilities and commercial stakes change. Neither position resolves without a shared, auditable record.
The technology itself will keep improving. That long arc depends on safety researchers being able to raise concerns internally and, where appropriate, externally, without breaking the controls that keep powerful systems and operational details contained. Getting that interface right is unglamorous work. It is also where credibility is built.


