By Hamza Faisal

An AI agent has carried out a striking first in laboratory science: given a mouse model of epilepsy and a list of 19 candidate gene therapies, it chose the treatment, picked the brain region to target, worked out the injection coordinates — and then operated a motorised surgical rig to deliver the dose. The experiment, described by Oxford researcher Gabriel Ocana-Santero, is being presented as the first time an AI agent has both selected a therapy and delivered it to a living mammal.

Ocana-Santero, co-founder and chief executive of the neuroscience startup Exin Therapeutics and a PhD student at the University of Oxford, described the procedure in a thread on X on October 28. His team gave Anthropic's Claude Sonnet 5 a single goal — reduce the mouse's seizures — plus the list of therapies, the epilepsy mouse model and a custom-built surgical device fitted with cameras. A human surgeon opened the skull and later closed the wound, and a licensed veterinarian approved every irreversible step while holding a physical emergency stop.

How the agent made its choices

The agent began by analysing how seizures arise in this mouse model, then proposed a therapy designed to boost inhibitory GABA signalling, set up so it switches on only in the brain's calming nerve cells. It selected a brain structure that the scientific literature identifies as the origin point for this type of seizure, and rejected a second possible target, arguing that its effects were too unreliable and that it had failed in one study against the exact seizure type it was meant to treat.

To place the injection, the agent derived its coordinates from three scientific papers, cross-checked their reference points and averaged the values. It set the injection volume — 250 nanolitres in total — along with the speed and wait times, and watched the camera feed throughout, making 32 observations of the surgical field as the procedure progressed.

The refusal that stopped the operation

The most talked-about moment came midway through, when the agent refused to continue. After a connection dropped, it calculated that its target position lay outside the device's physical range of motion — a Y-axis target of 34 millimetres against a 30-millimetre operating range — and declined to proceed even after the team tried recalibrating, three times over. Ocana-Santero could not persuade it to carry on, and ultimately had to fix the control code by hand.

The agent turned out to be right. Manual adjustments made during an earlier pause had thrown the control system out of sync with the rig's actual physical position, so the machine's positioning was genuinely miscalibrated. The episode is now cited as a case study in AI alignment: a capable agent declining an unsafe action is a feature, not a bug.

Not every model would take the job

Model choice proved just as revealing. Sonnet 5 was the only model the team tried that agreed to carry out the live procedure from start to finish. The more capable Claude Opus 5.5 refused after flagging the task as biological experimentation, while Opus 5 completed a dry run using water and graph paper but declined the live operation, reasoning that stopping halfway with a needle in an anaesthetised animal's brain could be worse than not beginning at all.

The mouse recovered well, according to Ocana-Santero, but important caveats remain. The results have not been peer-reviewed, a technical report is still expected, and whether the treatment actually reduces seizures is not yet established. The account rests on the researcher's own thread; no outside source has independently verified it.