Claude Sonnet 5 picks gene therapy, operates on mouse brain

The experiment sits at the intersection of two trends the account highlights: large language models gaining capability as control units for robots, and the alignment problem of making agents reliably follow human intent. Ocana-Santero describes the approach as "in vivo lab in the loop," connecting AI agents directly to experimental capabilities, and notes Anthropic is working on something similar with its "Model Hardware Standard."

The safety setup was layered. A licensed veterinarian approved every irreversible step and held a physical emergency stop, while fixed limits such as a cap on injection volume were enforced and the agent was forbidden from bypassing them. The mid-surgery refusal came after a dropped connection led the agent to detect its calculated target was outside the device's physical range of motion; restarting calibration did not change its assessment. A third attempt was made, and manual adjustments during an earlier pause had thrown off the control system, so the agent was right. Ocana-Santero ultimately fixed the code by hand.

The model choice itself reflected the alignment challenge: Sonnet 5 was the only model willing to carry out the live procedure from start to finish. Opus 5 said it could not guarantee completion and that stopping with a needle in an anesthetized animal's brain would be worse than not starting at all. Ocana-Santero notes that autonomous surgical robots already exist and are mechanically more precise than his lab device, but in previous cases a human always chose the surgery while the robot only carried it out. The mouse recovered well, according to Ocana-Santero.