Claude model autonomously performs mouse brain surgery, refuses to bypass safety limits

a claude model, specifically claude sonnet 5, was used to autonomously choose a gene therapy and perform surgery on a mouse brain. the model was given a list of 19 possible gene therapies, an epilepsy mouse model, and access to a custom motorized surgical device. it selected the drug, brain region, and stereotactic coordinates on its own, then controlled the procedure through the surgical rig. during the procedure, the agent made 32 observations and tracked the injection. when it detected a calibration error, it refused to continue, even when prompted to bypass safety limits. this incident highlights the ai alignment problem and the importance of ensuring ai agents reliably follow human intent.