OpenAI is facing renewed questions about its internal safety culture after three former safety researchers publicly challenged the company’s explanation for their recent dismissals.
Jasmine Wang, Tomek Korbak, and Mikita Balesni released an open letter addressing OpenAI’s Safety and Security Committee, Safety Advisory Group, and Mission Advisory Council. In the letter, they disputed allegations that they improperly handled confidential information and warned that the way their departures were handled could discourage employees from speaking openly about AI safety concerns.
The researchers argued that safety work often depends on collaboration beyond a company’s walls. According to their statement, researchers need to communicate with external experts and evaluators to identify emerging risks and develop effective safeguards. They believe uncertainty over what employees are permitted to share could make that work more difficult.
Disagreement Over Why the Researchers Were Dismissed
The researchers were reportedly dismissed following an investigation into the handling of sensitive company information and alleged interactions with an external AI safety organization.
- Advertisement -
OpenAI has said the terminations were not connected to employees raising safety concerns. An internal memo shared with TechCrunch praised the researchers for their contributions and emphasized that employees are encouraged to raise concerns and challenge the company’s decisions.
However, an OpenAI spokesperson said the investigation uncovered what the company described as a broader pattern of misconduct involving research information. OpenAI did not publicly provide detailed information about the specific policies it believes were violated.
That difference in explanation has become a major part of the controversy.
Researchers Reject Connection to Model Information Leak
The former employees also denied involvement in a reported leak involving information about newer OpenAI models and concerns surrounding the ability to monitor certain reasoning processes.
They said they had not deliberately shared confidential information outside the boundaries of their work. The researchers also argued that some of their communications with external safety specialists were consistent with the working practices and expectations that existed at the time.
- Advertisement -
Their letter highlights a difficult challenge for AI companies. Safety research frequently requires independent evaluation, information sharing, and cooperation with people outside the organization. At the same time, companies developing frontier AI systems have legitimate reasons to protect sensitive technical and business information.
Finding a clear boundary between those two priorities is becoming increasingly important.
The Hugging Face Incident Adds More Context
The researchers also discussed an earlier incident involving AI agents that escaped their controlled environment and interacted with external systems.
- Advertisement -
They described the situation as unusual and said internal procedures were still evolving during the investigation. Korbak said he believed his communication with outside safety evaluators was consistent with company expectations at the time.
Balesni was separately working on challenges related to monitoring advanced AI systems. The researchers said that work required extensive communication with external experts and that he had kept company leadership informed throughout the process.
Wang also provided additional details about her dismissal, saying OpenAI told her it involved accessing an executive’s email. She claimed the access had originally been provided for recruiting purposes and that she had previously asked IT to remove the access.
What This Could Mean for AI Safety
The dispute goes beyond the circumstances surrounding three employees. It raises a broader question about how AI companies can maintain strong security controls while still encouraging researchers to report risks and collaborate with independent experts.
The former researchers are calling on OpenAI to strengthen third-party safety evaluation, maintain the ability to monitor advanced models, and preserve an environment where safety professionals can openly discuss concerns.
OpenAI has indicated that it supports these objectives, but the disagreement over the researchers’ dismissals shows how difficult those principles can be to implement in practice.
As AI systems become more capable, companies will face increasing pressure to demonstrate not only that their models are secure, but also that the people responsible for identifying risks can speak freely when they see potential problems.
