Large models control robotic arms for surgery on live mice; founder rejects: who takes over after the robot stops?

Exin Therapeutics, a startup founded by Oxford neuroscientists, conducted an experiment where Anthropic's Sonnet 5 model autonomously planned and executed gene therapy injections into the brains of anesthetized epileptic mice. While the procedure was successful in terms of delivery, the AI model refused to move when controller coordinates exceeded safety limits and halted operations when interrupted, raising questions about fault tolerance and human oversight. The article analyzes the experiment's technical setup, compares it to surgical automation levels (L1-L5), discusses the limitations of current AI models in handling unexpected hardware states, and emphasizes the critical need for standardized interfaces and clear liability frameworks before such systems can be deployed beyond controlled laboratory settings.