An AI Agent Chose, Planned, and Performed Brain Surgery on a Live Mouse

Last week, Gabriel Ocana-Santero, co-founder and CEO of biotech startup Exin Therapeutics, conducted an experiment that sits somewhere between conventional biomedical research and science fiction: he gave an AI agent control of a motorized surgical arm, a mouse with an epilepsy model, and 19 gene therapy vectors, along with a single instruction—reduce seizures. The agent not only selected the treatment, the target brain region, and the injection coordinates, but operated the machine for more than an hour, monitoring the procedure in real time through a camera.

The experience, described by Ocana-Santero in a series of posts on the X platform on October 28, 2026, marks the first documented instance in which an AI system made a complete therapeutic decision and delivered it to a living mammal, without a human having chosen the protocol or the target.

Why gene therapy for epilepsy is so difficult

Epilepsy is one of the most complex neurological diseases to treat with gene therapies. Unlike conditions caused by a single defective gene—such as cystic fibrosis—epilepsy in laboratory models involves entire neural networks that fire excessively. To contain that electrical storm, researchers need something that strengthens the brain's inhibitory mechanisms without shutting it down entirely.

The agent was given 19 gene therapy vector options in a spreadsheet. Each vector carries a different gene, with distinct promoters and mechanisms of action. The wrong choice could have no effect or, in the worst case, exacerbate the seizures.

According to Ocana-Santero, the agent analyzed how seizures arise in this specific mouse model and proposed a therapy that increases GABAergic signaling—the brain's primary inhibitory neurotransmitter—under a promoter restricted to GABAergic cells. In practical terms, instead of silencing the entire brain, the therapy targeted specifically the hyperactive circuits that cause seizures, preserving the rest of neural function.

The agent also determined the brain region where the injection should be applied, cross-referencing information from three scientific papers to derive stereotactic coordinates. It calculated the volume (250 nanoliters), the speed, and the wait intervals—all without human intervention.

What the agent saw and how it reacted

During the procedure, the agent made 32 observations, according to Ocana-Santero. It monitored the surgical rig camera, tracking the injection and reporting whether there was any bleeding or backflow. At a crucial point in the surgery, the agent stopped.

There was a calibration error in the equipment. The agent calculated that its target position was outside the device's range and refused to continue, even when the team tried to recalibrate it. Later, the team discovered that the agent was right: previous manual adjustments had misaligned the control system. Ocana-Santero had to fix the code by hand to resume the procedure.

This moment of refusal—the agent saying "no" to a human request—is at once impressive and unsettling. It illustrates the alignment problem that AI developers have grappled with for years: how to make agents follow instructions without becoming so compliant that they ignore real risks.

The agent refused to continue because it had received the explicit instruction not to exceed its safety limits. When the connection dropped and it sensed the wrong position, it did not try to bypass the block. This suggests that the model maintains a functional sense of procedural integrity—something that, in the context of brain surgery, is a safety feature, not a bug.

Why Claude Sonnet, not Claude Opus

The choice of model also revealed a practical tension in the development of autonomous agents. The team initially tested Claude Opus 5.5—the most capable and most expensive model from Anthropic. It refused to guide the needle into the mouse's brain, arguing it might stop mid-surgery and endanger the animal.

The lighter model, Claude Sonnet 5, executed the procedure successfully. This raises an uncomfortable question: the "better" model is not necessarily the most suitable when autonomy is the goal. The analytical capability of Opus 5.5 came with a risk aversion that, in a surgical context, translated into an inability to act.

The procedure time was slightly longer than usual—just over an hour versus about 45 minutes—but the result was executed without complications. The mouse recovered well, and there are videos documenting the animal moving in its enclosure after the procedure.

What comes next

The therapy's effect on seizures has not yet been assessed. Researchers are waiting for the right moment to make that measurement. Ocana-Santero says a technical report will be published soon, but the results have not yet gone through peer review.

Exin Therapeutics is already testing other models and exploring whether AI-controlled agents can one day perform their own craniotomies—currently that step is still done by a veterinary surgeon. The company is also connecting agents to other types of experimental capabilities in the lab, which Ocana-Santero calls an "in vivo lab in the loop."

Anthropic, for its part, is working on a parallel project called the "Model Hardware Standard," which aims to standardize communication between AI models and physical hardware—the equivalent of the API that exists for software, but for the physical world.

The Exin Therapeutics experiment is a glimpse of what this could mean: not autonomous surgical robots in human operating rooms (something that is decades away, if it will ever be feasible), but AI agents making genuine therapeutic decisions in preclinical experiments. The boundary between "tool" and "participant" grows thinner with every step.

Sources: WIRED, The Decoder, Gabriel Ocana-Santero (X)

✓ Independent sources cross-checked and verified before publishing