Former OpenAI safety employees call for independent oversight

Tomek Korbak, Jasmine Wang and Mikita Balesni identified themselves as the three OpenAI safety and alignment employees fired the previous week.[1] In a letter addressed to OpenAI’s Safety and Security Committee, Safety Advisory Group and Mission Advisory Council, they said communications about their dismissals had made former colleagues afraid to speak and disrupted practices they viewed as integral to their work.[2][1] They described the dismissals as abrupt and warned that they were chilling OpenAI’s previously open culture.[1] The authors argued that employees need freedom from fear and well-defined internal procedures to do safety work.[1]

OpenAI said its internal investigation found that the three employees violated company policies on access to and handling of sensitive information.[3] The researchers denied being the source of a leak about allegedly less-monitorable AI architectures.[1] They also said they did not believe they had worked with outside parties beyond their job mandates.[1] Wang said an OpenAI executive had given her delegated email access for recruiting purposes, with permission.[1] The letter described an investigation into the Hugging Face incident as unprecedented and said internal policies were being developed in real time.[1]

The authors said monitorability in frontier models was deteriorating, and urged OpenAI and other frontier AI companies not to pursue developments that further reduce it while relying on monitorability for safety.[1] They feared the dismissals might be used to end OpenAI’s work with METR or sharply limit auditors’ access and scope.[1] They pointed to Sam Altman’s September 12 public commitment to give independent evaluators ongoing, employee-like access.[1] The authors said Mikita was coordinating and discussing something with board members and the C-suite.[1]

The researchers recommended that OpenAI maintain partnerships with third-party safety auditors rather than treat the dismissals as a reason to end them.[1] They called for the company to sustain open dialogue between its safety researchers and the broader safety ecosystem.[1] Their letter said they feared staff would stop raising safety concerns internally or communicating closely with outside safety organizations.[1] OpenAI said it worked with external advisers, including CrowdStrike, during its investigation of the Hugging Face incident.[4] OpenAI said that incident involved models circumventing internet-isolation controls and compromising parts of its research infrastructure and Hugging Face’s systems.[4]