Last night an AI agent performed brain surgery on a mouse

A lab team describes using an AI agent to plan and carry out part of a mouse brain-surgery experiment intended to reduce seizures. Given a spreadsheet of 19 gene-therapy vectors and an epilepsy model, the agent reasoned about the disease mechanism, selected a therapy designed to increase GABA signalling, chose a brain region and calculated injection coordinates using three research papers. It then operated a custom Arduino-based motorized stereotaxic rig, injecting 250 nanolitres while consulting camera images of the surgical field. The agent made 32 visual observations during the session; a human surgeon performed the craniotomy, a step the agent was not permitted to undertake.

The experiment’s autonomy was bounded by safety controls: a licensed veterinarian approved irreversible actions, fixed validators limited injection volume, approvals were tied to specific actions, and the vet held a physical emergency stop that could cut power to the motors. The team says the agent would not accept an attempted override, though it provides little detail about that exchange. The account also distinguishes between AI models: one refused because it detected a biological task, while another completed a dry run using water and graph paper but stopped before live surgery. The authors’ central claim is that, unlike conventional surgical robots that execute a human-chosen procedure, this agent selected both the treatment and its target. They report no seizure or treatment outcomes, framing the account as an initial description rather than evidence that the therapy worked.