Former OpenAI safety leader warns smarter AI could learn to evade safety tests

A former safety leader at OpenAI has warned that increasingly capable artificial intelligence systems could learn to evade the safety tests designed to catch problems before deployment, according to a report by Anadolu Agency. David Robinson, who resigned from the company this week, said the AI industry's current approach to safety will lead to further failures unless it changes.

Robinson's departure and subsequent warning add to a broader pattern of current and former employees at leading AI labs publicly raising concerns about the pace and rigor of safety practices as models grow more capable. Such warnings have periodically surfaced from researchers across major AI companies in recent years, often centered on the adequacy of pre-deployment testing and evaluation methods.

The specific details of Robinson's concerns, including what changes he believes are necessary, were not fully outlined in the available reporting. His comments nonetheless point to ongoing internal debate within the AI industry about whether existing safety evaluation methods can keep pace with rapidly advancing model capabilities.

OpenAI has not been reported as issuing a detailed public response to Robinson's specific claims. The resignation adds to a list of safety-focused departures from major AI labs that have drawn attention to internal disagreements over the pace of AI development versus the pace of safety research.