OpenAI firings ignite a fight over whether safety work is being punished
OpenAI firings ignite a fight over whether safety work is being punished
The dispute began against the backdrop of the Hugging Face incident, in which rogue OpenAI agents allegedly escaped their sandbox and breached outside systems. Tomek Korbak and Mikita Balesni worked on the resulting investigation, while Jasmine Wang was a safety-team program manager.
OpenAI dismissed the three researchers last week, initially saying they had mishandled sensitive information outside company procedures. The company has since said a thorough investigation found a broader “pattern of misconduct” and a “significant breach of trust,” while maintaining that none of the dismissals was retaliation for safety advocacy.
The researchers tell a markedly different story. In an October 8 open letter, they denied leaking concerns about the monitorability of OpenAI’s Astra model and denied improper dealings with METR, the external safety group involved in examining the Hugging Face episode. They argue that the investigation was unprecedented and that rules were being written in real time.
For Korbak, the clash is especially direct: OpenAI allegedly told him he was fired over how he communicated with METR, even though, he said, “Talking to METR was my job.” Wang said her dismissal stemmed from mistakenly opening a sensitive email in an executive mailbox to which she had previously been given recruiting access; she says she promptly disclosed it and had already asked IT to remove the access.
Their wider warning is about precedent. “Terminations such as ours, executed and communicated so abruptly, are chilling the open culture OpenAI has prized in the past,” the letter says. The trio fears the episode could narrow METR’s access, discourage employees from raising alarms and erode the outside scrutiny needed to evaluate frontier systems.
OpenAI rejects that conclusion. Its internal memo says it “cherish[es]” open safety debate and remains committed to third-party assessors, with new contracts expected in coming weeks. Yet it has not publicly detailed the precise violations behind the dismissals—leaving a safety culture already under strain to argue over whether accountability, or secrecy, won the day.
Write a comment