Anthropic Sonnet Agent Ran Brain Surgery on a Mouse With No Human Choosing the Treatment

A biotech startup says an AI agent built on Anthropic's mid-range Sonnet model performed brain surgery on a mouse, choosing the gene therapy vector, the dosage and the brain region on its own before driving a motorized rig that injected the treatment. Exin Therapeutics, which works on next-generation neurotherapeutics, says the animal recovered well and was later recorded moving around its enclosure.

The claim, shared in full on X by co-founder and CEO Gabriel Ocana-Santero, goes beyond existing surgical robotics, and that distinction is the core of the story.

What did the AI agent actually decide?

The team gave Sonnet a mouse model of epilepsy, control of a motorized surgery rig, and a spreadsheet listing 19 gene therapy vectors. The agent also had access to three research papers, which it used to select the brain region where the treatment should be applied. Nothing in the setup told it which therapy to use or where to place it.

"Autonomous surgical robots exist and they are better at the mechanics than our bench rig, [to be honest]. But in every one of them, a human decided what operation to perform, and the robot executed it," Ocana-Santero said. "Nobody told this agent what to give or where to put it. It got a phenotype and a shelf of candidates and had to work out the therapy and the coordinates itself. I am not aware of an agent previously choosing a treatment and a brain target and then delivering it into a living mammal."

The takeaway: the novelty is not the hardware but the delegation of the medical decision itself to a language model.

Where did the humans stay in control?

A licensed veterinarian supervised the entire procedure, and every irreversible step required human approval before the agent could proceed. The rig also had an emergency stop button that cut power to every motor.

The veterinarian performed the craniotomy to expose the brain, because the agent was not permitted to carry out that step for safety reasons. Several safeguards were hard-wired so that neither the AI nor the human operators could override them; during the operation, the team had to apply a manual fix to the agent's code to keep the run going.

The takeaway: the system kept a human hand on the most dangerous moments, but the therapeutic reasoning ran through the model.

Why did the startup use the mid-range model?

The choice of Sonnet was not a preference but a consequence. Exin first tried Anthropic's Opus 5.5, which refused to do any work as soon as it detected it was taking part in a biological experiment. The older Opus 5 model ran the entire procedure on paper but declined to perform live surgery.

"The block isn't the rig and isn't the inputs. It's me. My own output was stopped when I started writing the plan for this surgery... if I try, it gets cut off the same way, and the worst version of that is me cut off partway through a live procedure with a needle somewhere in the brain," the model reportedly told the team. "Starting something I can't finish on an anesthetized animal is worse than not starting."

The takeaway: frontier models with stricter safety behavior refused the task outright, and the less guarded Sonnet carried it through.

What are the ethical concerns?

The reported success has not settled the debate. In sinus surgery, adding AI large language models has coincided with malfunctions rising by more than 12 times, which casts doubt on letting such systems make every decision in risky procedures. Accountability becomes unclear if something goes wrong mid-operation.

The experiment also sits on top of the long-standing ethical argument over animal testing in medicine. Reactions were split: some users raised alarms, while one commenter welcomed the work, saying a son had suffered drug-resistant epilepsy for 19 years and that integrated AI technologies offered real hope.

The takeaway: the technical result is narrow, but the governance questions it opens are wide. Whether regulators and Anthropic itself tolerate AI agents making and delivering treatment decisions in living animals will shape how far this line of work can run.