An AI agent called Claude Sonnet 5 selected a gene therapy, chose where to inject it in a mouse’s brain, and operated a motorized surgical device during the procedure. The mouse had an induced-seizure model of epilepsy. The agent’s instruction was to reduce the seizures, according to Gabriel Ocana-Santero, an Exin Therapeutics cofounder and University of Oxford PhD student who described the experiment on X.
The team gave the agent a choice of 19 gene therapies. It analyzed how seizures arise in this mouse model, selected a therapy intended to boost inhibitory signaling in the brain, and picked a brain structure that research identifies as a source of this type of seizure. It rejected another possible target, citing concerns about unreliable effects and a study in which that target had failed against the same seizure type.
What happened during the procedure?
The agent used three scientific papers to work out injection coordinates, cross-checked their reference points, and averaged the values. It also set the injection volume, speed, and wait times. During surgery, it watched a camera feed and made 32 observations, including whether it saw bleeding, Ocana-Santero says.
People remained involved in safety. A licensed veterinarian approved each irreversible step, and a physical emergency stop could cut power to the device’s motors. The system also had fixed limits, including a cap on injection volume that the agent could not bypass.
One safeguard stopped the procedure. After a connection dropped, the agent calculated that its target position was outside the device’s range and refused to continue, even after the team tried recalibrating it. The agent was right: manual adjustments earlier had disrupted the control system. Ocana-Santero ultimately had to fix the code by hand.
A demonstration, not a clinical result
Ocana-Santero says he is not aware of an earlier case in which an AI agent both chose a treatment and brain target and delivered it to a living mammal. He contrasts that with surgical robots in which a human chooses the operation and the robot carries it out. The mouse recovered well, he says, but the experiment’s results have not been peer-reviewed; a technical report is expected.
The choice of model also exposed a practical tension. Sonnet 5 was the only model the team tried that agreed to carry out the live procedure from start to finish. Claude Opus 5.5 and Opus 5 refused. Opus 5 said it could not guarantee completing the operation and that stopping with a needle in the anesthetized animal’s brain could be worse than not beginning. The episode shows both how much responsibility an AI system can take on in a lab and how a refusal can become consequential once a procedure is underway.
Why it matters
For laboratory researchers, the demonstration suggests an AI agent can make treatment and targeting choices as well as operate lab equipment, rather than merely follow a human’s surgical plan. But the agent’s refusal also halted a time-sensitive procedure, and the account does not establish that the treatment reduced seizures or that this setup is safe or reliable beyond this experiment.
