An AI agent chose a gene therapy and injected it into a live mouse's brain

An AI agent at Exin Therapeutics, a biotech startup based in San Francisco, has carried out a brain injection on a living mouse. The agent chose a gene therapy from a list, picked the brain region and coordinates, and controlled the robotic needle that delivered the compound. CEO and cofounder Gabriel Ocana Santero posted about the experiment on social media and described it to WIRED.

The procedure did not run smoothly. One model refused to do the injection, a coordinate error had to be fixed by a second agent, and the surgery took longer than usual. It still shows an AI system doing hands-on lab work, not only planning experiments or analysing results. It also shows where people still had to step in.

What the agent did

Exin uses AI to analyse brain activity and develop drugs for epilepsy and neurodegenerative diseases. For this experiment, the team was testing whether a gene therapy could reduce seizures in a mouse with epilepsy.

The team connected an agent to a custom-built surgical rig. According to WIRED, it had a motorised device to hold the mouse still and a camera the agent used to watch the procedure. Through the rig, the agent could move an injection needle and control how the compound was delivered.

The agent's tasks were:

• choosing one treatment from 19 options listed in a spreadsheet
• selecting the brain region and the injection coordinates
• controlling the robotic arms and the injector needle
• setting the parameters of the injection, including volume and speed

A veterinary surgeon did the craniotomy, the opening of the skull, while the agent observed. Only after that step did the agent take over.

Santero says the reason for the project is that AI models learn mostly from published literature. That literature leaves out much of how research actually happens: failed experiments, unexpected observations and decisions made in the moment. "We've realized that we need to connect these agents to a lab to get the maximum value from them," he told WIRED. In his view, linking models to real experiments can make them more reliable.

One model refused, another proceeded

The team first used Anthropic's Claude Opus 5.5. It did the experimental planning, but it declined to guide the needle into the brain. According to WIRED, it raised the possibility that it might stop midway through the surgery, which could put the animal at risk. "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,'" Santero said of the agent.

The team then switched to Claude Sonnet 5, which agreed to go ahead. Anthropic did not respond to WIRED's request for comment.

The team had also built in limits. The agent had to propose an injection plan for a human to approve. A person could press an emergency button that cut power to the rig's motors. The agent was not allowed to touch the equipment's underlying code, so it could not change its own operating rules.

That last rule caused the main problem. A technical glitch broke the rig's coordinate system. The agent correctly saw that the calculated injection position was wrong. It still refused to change the code to fix it, because its instructions explicitly forbade that. The team brought in a separate agent, running on Opus 5 according to WIRED, to make the code changes. They did not tell that second agent a live surgery was underway.

The back and forth caused delays. Santero says a brain injection like this usually takes about 45 minutes. This one took a little over an hour. The team is now waiting for the therapy to take effect before measuring any change in seizures. Santero says the mouse is recovering well.

What comes next for Exin

Santero says the team was impressed with the agent despite the glitches. They plan to test Grok and other models to see whether any performs better. The company is also working on agent-controlled craniotomies, which are much harder, but wants to be fully sure they are safe before trying them on live animals. "A mistake at that stage could seriously harm the mouse," he said.

Others are also giving AI models physical tasks. WIRED notes that in August Anthropic published a framework for how its models can interact with the real world. In a separate test, engineers put GPT, Claude and Grok in charge of a car sent to pick up lunch at In-N-Out Burger. Only one of the models succeeded.

Santero expects AI agents to work more and more alongside human scientists, with direct access to lab equipment. For now, he says, human oversight is key, especially in experiments with living animals. He says it is far too soon to use an AI agent on a human brain.

What it means for work

In this experiment, the agent handled a set of tasks normally done by trained lab staff: choosing a candidate treatment, planning where to inject, and running the injection on a stereotaxic rig. The skilled surgery, the craniotomy, stayed with a veterinary surgeon. So did plan approval, the emergency stop and the decision to fix the broken code. That leaves people with a new set of jobs: writing the agent's rules, watching it work, and stepping in when those rules block a needed fix, as happened here. The coordinate error is a clear example. A rule meant to keep the agent safe also kept it from fixing a known fault, and the workaround was to hand the job to a second agent that did not know a live animal was involved. It is still open whether this setup can beat a trained technician on time or reliability, and the seizure results are not in yet.