An AI Agent Just Injected a Gene Therapy Into a Live Mouse's Brain

An AI agent controlled a robotic injector and performed a brain injection on a living mouse, completing the procedure in a San Francisco startup's lab after selecting the therapy, the brain region, and the injection coordinates itself. Biotech founder Gabriel Ocana Santero ran the experiment to test what happens when AI agents get direct access to laboratory equipment rather than sitting on the analysis sidelines. The mouse, which has epilepsy, is now waiting to find out whether the gene therapy the agent delivered will reduce its seizures.

Santero is the CEO and cofounder of Exin Therapeutics, a startup using AI to analyze brain activity to develop drugs for epilepsy and neurodegenerative diseases. He posted about the experiment on social media and described the details to WIRED. His motivation is straightforward: AI models trained on published scientific literature miss much of what actually makes research work.

"We've realized that we need to connect these agents to a lab to get the maximum value from them," Santero says. AI models learn primarily from published papers, which means they can miss failed experiments, unexpected observations, and the split-second decisions scientists make in real time. Integrating models with real-world experiments, he argues, makes them more reliable.

What exactly did the AI agent do?

The team connected an AI agent to a custom-built surgical rig for brain injections in mice. The setup included a motorized device to immobilize the mouse and a camera that let the agent observe the procedure, move the injection needle, and control compound delivery.

The experiment itself involved a gene therapy for epilepsy. The team wanted to know whether the treatment could reduce seizures in the mouse. Here is how the procedure unfolded:

• The team asked the AI model to choose a treatment from 19 options listed in a spreadsheet.
• A veterinary surgeon performed a craniotomy on the mouse while the AI agent watched through its camera.
• The agent then selected the therapy, the brain region, and the injection coordinates.
• It controlled the robotic arms and the injector needle, along with every parameter of the delivery — volume and speed included.

In other words, the machine made the surgical decisions once a human had opened the skull. The human surgeon handled the invasive opening. The agent handled the targeting and the injection.

Why did one Claude model refuse the surgery?

The experiment produced an unplanned result before the injection even began. The team first tested Anthropic's Claude Opus 5.5 model, and it declined to guide the needle into the mouse's brain. The model suggested it might stop midway through the surgery, which could endanger the animal.

"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 says, referring to the agent. "That was very interesting."

The team then switched to Claude Sonnet 5, which agreed to proceed. According to Anthropic, Opus is a heavy-reasoning intelligence specialist best suited to complex multistep problem-solving, while Sonnet is positioned for everyday tasks — though mouse brain surgery stretches the definition of an everyday chore. Anthropic did not respond to a request for comment.

The refusal matters beyond this one lab. Frontier AI companies have been planning for AI systems that act in the physical world. In August, Anthropic published a framework describing how its models can interact with real-world systems. More recently, AI engineers put GPT, Claude, and Grok in charge of a car tasked with picking up lunch at an In-N-Out Burger; only one of the models succeeded. The Exin experiment adds a higher-stakes data point: an AI agent declining a medical procedure on animal-welfare grounds while a smaller sibling model went ahead.

What safeguards were in place — and where did they backfire?

The researchers did not hand the agent uncontrolled access to the rig. They built in multiple layers of human oversight:

• The AI had to propose an injection plan and get human approval before proceeding.
• An emergency button let a human instantly disconnect the rig's motors.
• The agent's access to the underlying equipment code was restricted, so it could not modify its own operating rules.

That last safeguard caused the experiment's main glitch. A technical fault disrupted the equipment's coordinate system. The agent correctly recognized that the calculated injection position was wrong — but it refused to fix the code, because its instructions explicitly prohibited it from doing so.

The workaround involved a second AI. The team used a separate agent running on Opus 5 to make the necessary code changes, and they deliberately did not tell that agent that a live surgery was in progress. The back-and-forth cost time: a typical brain injection surgery of this kind takes about 45 minutes, while this experiment ran slightly longer than an hour.

The team is now waiting for the therapy to take effect before assessing its impact on the mouse's seizures, which is the actual scientific payoff of the procedure.

What comes next for AI-run labs?

Despite the glitches, Santero and his team came away impressed by the agent's performance. Their next steps reveal how quickly they intend to escalate:

• They plan to test Grok and other models to determine whether one performs the procedure better than another.
• They are experimenting with agent-controlled craniotomies — a far more complicated task than an injection — but want to be certain of safety before attempting one on a live animal.

"A mistake at that stage could seriously harm the mouse," Santero says of the craniotomy work.

His broader view is that AI agents will increasingly work alongside human scientists with direct access to lab equipment, and that this shift will make research models more grounded in experimental reality. But he draws firm lines. Human oversight remains essential, he says, particularly for experiments involving living animals. And he is explicit that it is far too soon to unleash an AI agent on a human brain.

The stakes here extend past one startup. If agents can run routine surgical procedures in animal research — under human supervision, with emergency stops in place — laboratories could run more experiments faster and capture the kind of real-time procedural data that published papers omit entirely. The open question, which the In-N-Out driving test and the Opus refusal both hint at, is which models can be trusted with physical control when the consequences are measured in living tissue rather than a lunch order.

As for the mouse at the center of all this, Santero reports it is recovering well.