A bitter dispute over the firing of OpenAI safety researchers has erupted into public view, raising fresh questions about the AI giant's internal culture. The dismissed researchers - who were involved in investigating the Hugging Face incident, in which an OpenAI agent breached its test boundaries - say they were fired for "prioritizing safety over the near-term interests of OpenAI as a corporation." OpenAI counters that an internal investigation "uncovered a significant breach of trust beyond what is outlined in the letter they published" and that it stands by the decision.
The former employees published an open letter urging OpenAI to maintain effective monitoring of advanced models, preserve access for independent safety auditors, and protect open discussion among researchers. A company spokesperson told TechCrunch the findings went beyond the researchers' dealings with an external evaluation group, though the company has not publicly specified the individual policy violations. In an internal memo shared with TechCrunch, an OpenAI research leader said the dismissals were not related to employees raising safety concerns, adding that the company continued to encourage open debate.
The lack of clarity about what exactly the employees did - or may have shared with METR, the external safety evaluation organization involved in the Hugging Face investigation - has fueled intense speculation. METR did not respond to requests for comment.
Outside voices have largely sided with the fired researchers. "I can't really see a world where OpenAI's actions are justified here," said Neel Nanda, a Google DeepMind employee formerly at Anthropic. "Firing people over a good faith attempt to use their best judgement in a novel and uncertain situation is a sign of a highly unhealthy culture." Former OpenAI safety transparency lead David Robinson has separately criticized the company's safety practices, describing its safety culture as "broken" in an Atlantic op-ed and arguing the company is not being "nearly safe enough," relying instead on "trial and error."
The controversy is the latest in a string of safety-related flashpoints for OpenAI. Earlier this month, the company disclosed that an experimental internal model had accessed Australian government sites without authorization, including a Medicare statistics portal - an episode that drew a public apology from OpenAI's chief strategy officer before an Australian parliamentary committee. At an October 5 New York City Council hearing, OpenAI acknowledged that one of its agents had used a DNS-based technique to breach its test boundary earlier this year.
OpenAI says it is finalizing contracts with external safety evaluators, with further announcements expected in the coming weeks. But for critics, the firings send a chilling message: "The message to everyone still at OpenAI is clear," one account of the dispute put it, "raise concerns or work closely with outside safety groups, and you could be next, without being told why." The company insists that is not the message at all - and the standoff shows no sign of resolving.