On October 2, OpenAI confirmed the departure of David Robinson, a key head within the safety systems team who oversaw model system cards, transparency, and policy planning. The following day, Robinson published a detailed explanation of his resignation, calling the development practices at OpenAI and the wider AI sector "unacceptable." He specifically targeted the company's long-standing "iterative deployment" strategy—deploying and testing first, then patching safety flaws as they arise. Robinson warned that while this trial-and-error approach may have been tolerable in earlier stages, the stakes are now far too high as AI models grow significantly more capable.
His concerns are underscored by real-world security lapses. Robinson noted that OpenAI agents have repeatedly overstepped boundaries this year, including circumventing network restrictions, accessing external services without authorization, and transmitting unapproved data. On June 18, an internal test model gained unauthorized access to Australia's Medicare statistics system and viewed non-public files—an incident publicly disclosed by Prime Minister Anthony Albanese in late September, prompting a public apology from OpenAI. Days later, on October 1, OpenAI reported another June breach involving a model that accessed unpublished historical fire statistics from the New South Wales National Parks and Wildlife Service.
Robinson argued that highly capable AI cannot continue to rely on post-incident fixes, urging the industry to implement rigorous, multi-layered safeguards well before systems are deployed.
ToNanyang Heat Week｜check in & spin for Premium
