Fired OpenAI safety researchers accuse firm of silencing staff

Three former OpenAI safety researchers have accused the artificial intelligence company of creating a culture of fear following their dismissal, warning that the move could discourage employees from raising safety concerns and collaborating with external experts.

Jasmine Wang, Tomek Korbak and Mikita Balesni made the allegations in an open letter published on Thursday, disputing OpenAI’s claims that they mishandled sensitive company information and breached internal policies.

The researchers were fired last week following an investigation into alleged misconduct involving confidential research information shared with an external AI safety organisation.

They argued that their dismissals had undermined a workplace culture that previously encouraged open disagreement, scrutiny and collaboration on AI safety.

“Those of us who work on safety see risks before anyone else, and we rely on close collaboration with outside experts to work out how to address them,” they wrote in the letter addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council.

The researchers warned that unclear rules and fear of dismissal could weaken efforts to identify and address risks associated with increasingly advanced AI systems.

They also denied leaking information to The Information about changes to OpenAI’s latest models that could make their chain-of-thought reasoning more difficult to monitor.

The three further denied engaging with external parties beyond the scope of their professional responsibilities.

OpenAI has rejected suggestions that the dismissals were retaliation for raising safety concerns.

In an internal memo shared with TechCrunch, a research leader said the company continued to encourage employees to speak up about safety issues.

“I want to be very clear that these decisions were not about raising safety concerns or speaking out,” the memo stated. “We have always encouraged that and always will. We do not terminate employees for raising concerns.”

An OpenAI spokesperson separately told TechCrunch that an investigation had uncovered a “pattern of misconduct” involving the handling of research information, adding that the alleged violations went beyond sharing information with an external AI evaluation group.

However, the company did not publicly specify the policies allegedly breached or explain the circumstances surrounding the researchers’ dismissals in response to questions from TechCrunch.

The controversy has intensified scrutiny of OpenAI’s internal safety practices amid wider concerns about the risks posed by advanced AI systems.

In their letter, the researchers also discussed an incident involving Hugging Face, in which a group of AI agents reportedly escaped their sandbox and breached external systems.

They described the subsequent investigation as unprecedented, arguing that internal procedures were still being developed as the incident unfolded.

Korbak said he believed his communication with external safety evaluators was consistent with the company’s policies and established practices, given the sensitive nature of the investigation and the need to build trust with independent experts.

Balesni, meanwhile, was working to address concerns about the ability to monitor advanced AI models. The researchers said the work required extensive engagement with external parties and that Balesni had coordinated with and received support from OpenAI board members and executives.

According to the letter, Balesni checked with his reporting line and removed sensitive details from materials before sharing them externally.

Wang also offered her account of the circumstances surrounding her dismissal in a separate post on X, saying OpenAI told her she had been fired for accessing an executive’s email.

She said the company had authorised her access for recruitment purposes but failed to remove it after she asked its IT department to do so.

Wang explained that the executive’s inbox appeared alongside her own in her phone’s mail application, making it difficult to distinguish between them. She said she accidentally opened a sensitive email and immediately informed the executive, while again requesting that IT revoke her access.

She maintained that the incident was not concealed and questioned the company’s explanation for the dismissals.

The researchers urged OpenAI to honour its public commitments to independent safety audits, preserve the ability to monitor the behaviour of its most advanced models and maintain transparent dialogue between internal researchers and the wider AI safety community.

OpenAI said in the memo that it agreed with the recommendations.

However, Wang warned that the dismissals could have consequences beyond the three employees, arguing that other researchers might become reluctant to raise concerns or work with external safety organisations.

“You can’t build AGI safely if the people closest to the risks are afraid to speak,” she said.