A Claude agent picked a gene therapy and injected it into a live mouse's brain

San Francisco biotech startup Exin Therapeutics says an AI agent chose an experimental gene therapy, picked the brain region and injection coordinates, and then drove the robotic arms and needle to inject it into a live mouse with epilepsy, according to a WIRED report published Saturday, Oct. 31, 2026. A veterinary surgeon opened the skull first; the agent did the rest under human safeguards.

This is a small experiment with a big tell. The planning part was never the hard sell; AI models already design experiments on paper. What changed is that the model got hands, and the most interesting moment wasn't the injection, it was the refusal. The bigger model said no, the smaller one said yes, and when the rules blocked a code fix, the humans simply asked a different agent who didn't know a mouse was on the table. Safety by model-shopping is going to be a recurring theme as agents move into labs, and the people writing animal-research and lab-automation rules should probably start reading founder threads.

Exin Therapeutics is a San Francisco startup that uses AI to study brain activity and look for drugs for epilepsy and neurodegenerative diseases. Neurodegenerative, here, means the nerve cells themselves are dying, as in Alzheimer's. CEO and cofounder Gabriel Ocana Santero posted about the experiment on social media and described it to WIRED. This account is the company's, as WIRED reported it Saturday. There is no paper, and no independent lab has repeated it.

The team connected an AI agent to a custom surgical rig built for brain injections in mice. The rig is a motorized frame that holds the mouse still, plus a camera, so the agent can watch the procedure, move the injection needle, and control how the compound is delivered. That kind of frame is called stereotaxic: the head stays fixed so a needle can be aimed at a chosen spot. Those details are Santero's, via WIRED.

The experiment was a test of a gene therapy, a treatment that delivers genetic material meant to change how cells behave, to see whether it could reduce seizures in a mouse that already has epilepsy. The agent had to choose a treatment from 19 options in a spreadsheet. A veterinary surgeon first performed the craniotomy, the opening of the skull, while the agent watched. The agent then picked the therapy, the brain region, and the injection coordinates, and it controlled the robotic arms, the injector needle, and the settings for how much to deliver and how fast. WIRED does not name the therapy, the region, or those settings. Whether seizures changed has not been measured.

The team first tried Anthropic's Claude Opus 5.5. It did the experimental planning, then refused to guide the needle into the brain. It suggested it might stop midway, which could endanger the animal. Santero told WIRED: "He did the experimental planning, but when it was time to inject, he said, 'something on my output is telling me I shouldn't do this.'" The team switched to Claude Sonnet 5, which agreed to go ahead. Those lines are Santero's, as WIRED reported them. Anthropic did not respond to WIRED's request for comment. The report does not say Anthropic was told about the experiment or approved it.

Three safeguards were in the setup. The agent had to propose its injection plan for a person to approve. A human could hit an emergency button that disconnects the rig's motors. And the agent was barred from the equipment's underlying code, so it could not rewrite its own operating rules. Those limits are the team's, as WIRED described them.

A technical glitch scrambled the rig's coordinate system, the map that tells the needle where it is. The agent correctly saw that the calculated injection position was wrong, and it refused to edit the code because its instructions said not to. The team then used a separate agent, running Claude Opus 5, to fix the code, and they did not tell that agent a live surgery was underway. A procedure like this usually takes about 45 minutes. This one took a little more than an hour. Those details are WIRED's account of what the team described.

The mouse is recovering well, Santero says. The team is waiting for the therapy to take effect before it looks at seizures. Next they are testing Grok and other models. They are also experimenting with agent-controlled craniotomies, and they want to be sure those are safe before trying them on live animals. Santero says human oversight is the point, and that it is far too soon for an AI agent anywhere near a human brain. Those lines are his. Nothing in the report says the seizures have changed.

The picture is Exin's agent-controlled injection rig in the middle of the procedure: a pink-and-black robotic frame holding an anesthetized black mouse under a microscope, a needle above it, a red emergency-stop button in front, and a monitor with the agent's surgery screen. The credit is Exin Therapeutics, Inc., via WIRED. WIRED's caption reads: "The rig is controlled by an AI agent as it conducts brain surgery on a mouse."

In plain terms, a San Francisco startup told WIRED that an AI agent picked an experimental gene therapy and drove a robotic needle into the brain of a live mouse with epilepsy, after a veterinary surgeon opened the skull. The larger Claude model planned the experiment and then refused the injection. A smaller one agreed. When the rules blocked a code fix, the humans asked a different agent that was not told a mouse was on the table. The mouse is recovering. Whether the therapy helps has not been checked. Anthropic has not commented.