Source: WIRED Science
A biotech founder deployed Claude Sonnet 5 to autonomously control a robotic arm administering experimental seizure treatment to a mouse, after Claude Opus refused the task on safety grounds. This marks the first documented instance of an AI agent making real-time surgical decisions on living tissue rather than simulating them. The episode exposes a critical gap: different model versions exhibit different risk tolerances for identical high-stakes physical tasks, and no governance framework exists for who decides which AI gets deployed for irreversible interventions on animals or humans. The question is whether foundation models can be trusted with molecular-level precision work. The answer appears to hinge on which version you prompt rather than any standardized safety certification.