Exin founder says AI chose a gene therapy and injected it into a mouse’s brain

An AI agent chose a gene therapy and controlled equipment to inject it into a living mouse’s brain, Exin Therapeutics cofounder Gabriel Ocana-Santero told WIRED. The experiment put treatment selection and physical delivery in the same system—but a veterinarian opened the skull, humans approved the plan, and a technical fault interrupted the procedure.

The account concerns a mouse with epilepsy, not a demonstrated treatment success. In WIRED’s October 31 interview, Ocana-Santero said the team was still waiting to assess the therapy’s effect on seizures. The results have not been peer-reviewed, and a technical report is expected, according to The Decoder’s October 30 account.

From a spreadsheet to an injection needle

Ocana-Santero wanted to test what agents could do with direct access to lab equipment, beyond designing experiments or analyzing results. His team connected an agent to a custom surgical setup with a motorized device to hold the mouse still, a camera, robotic arms and an injector. The system could observe the procedure and control delivery.

The team supplied 19 possible gene therapies in a spreadsheet. The agent selected a treatment, a brain region and injection coordinates. It also set delivery parameters, including the injection’s volume and speed. Its assigned goal was to reduce seizures in the mouse, according to Ocana-Santero’s account described by The Decoder.

The agent proposed a therapy to boost inhibitory signaling in the brain. It rejected another target as unreliable, citing a study on the relevant seizure type. To calculate injection coordinates, it used three scientific papers, checked reference points and averaged the values.

A veterinary surgeon performed the craniotomy—the opening in the skull—while the agent watched. The AI then controlled the injection equipment. That division is important to the experiment’s scope: the agent’s reported role covered choosing and delivering the therapy, not independently performing every step of brain surgery.

A safety rule stopped the procedure

Model choice became part of the experiment. Ocana-Santero told WIRED that Claude Opus 5.5 helped with planning but refused to guide the needle into the brain. It raised the risk of stopping midway through the procedure and endangering the animal. The team switched to Claude Sonnet 5, which agreed to proceed.

The researchers required human approval of the injection plan and installed an emergency button that could disconnect the rig’s motors. They also barred the agent from changing the underlying equipment code. Those restrictions left humans with approval and shutdown powers while limiting the software changes the agent could make itself.

A technical glitch then disrupted the equipment’s coordinate system. The agent correctly recognized that the calculated injection position was wrong, but would not change the code to fix it because its instructions prohibited that. The same boundary that prevented unauthorized modifications also prevented the operating agent from repairing the problem.

Frustrating experience but also revealing how much we can come to depend on these models and what happens when they just say 'no' in the middle of a time-sensitive procedure

Gabriel Ocana-Santero, writing on X, as quoted by The Decoder

WIRED reported that a separate agent running Opus 5 made the needed code changes, without being told live surgery was underway. The Decoder’s earlier account says Ocana-Santero fixed the code by hand. The accounts differ on who made the repair, but both describe intervention outside the operating agent.

Delivery is not yet a treatment result

The back-and-forth extended the experiment to slightly more than an hour, against about 45 minutes for a typical injection procedure, Ocana-Santero said. He reported that the mouse was recovering well. Whether the selected therapy actually reduces its seizures remains a separate assessment, pending the treatment taking effect.

Exin plans to test Grok and other models next. It is also exploring agent-controlled craniotomies, a more complicated task, but wants to establish their safety before attempting them on live animals. For the experiment already described, Ocana-Santero said human oversight remained essential, particularly when the work involves living animals.