AI Brain Surgery: Claude Agent Injects Gene Therapy Into a Mouse

An AI agent built on Anthropic’s Claude Sonnet 5 chose a gene therapy, picked a brain target and injected the treatment into a living mouse at San Francisco startup Exin Therapeutics, the company’s CEO Gabriel Ocana-Santero said. Ocana-Santero described the AI brain surgery on X on October 28 and said the mouse recovered well. Whether the therapy actually reduces seizures is not known yet.

Last night an AI agent performed brain surgery on a mouse.

We gave it 19 gene therapy vectors in a spreadsheet, a mouse model of epilepsy, access to a motorize surgery rig with cameras and one goal: reduce seizures.

It picked the drug. It picked the brain region. It derived the coordinates from three papers.

— Gabriel Ocana-Santero (@OcanaSantero) October 28, 2026

How the AI brain surgery worked

Exin gave the agent a spreadsheet of 19 gene therapy vectors, a mouse model of epilepsy, control of a custom motorized stereotaxic rig built on Arduino and a single goal: reduce tonic-clonic seizures. The agent settled on a therapy that boosts GABA, the brain’s main inhibitory chemical messenger, in neurons that use it. It placed the injection in the lateral entorhinal cortex, where this type of seizure is reported to start.

The coordinates were averaged from three published papers: 3.90 mm behind bregma, 4.00 mm to the right and 4.30 mm deep. The agent then drove the needle and injected 250 nanoliters, about 500 million viral genomes, over five minutes. It checked the rig’s camera 32 times during the session, at one point reporting “37.5 of 250 nL so far, about 15 percent. No bleeding or backflow visible at the opening.”

Humans kept control of the riskiest steps. A licensed veterinary surgeon performed the craniotomy, which the agent is not yet allowed to do, approved every irreversible step and held a physical emergency stop that cuts power to every motor, Ocana-Santero wrote. The setup echoes the wider debate over an AI kill switch for powerful models.

Why two Claude models said no

Claude Opus 5.5 refused immediately once it detected a biology task, according to the CEO. Claude Opus 5 completed a full dry run on water and graph paper but stopped before the live procedure, saying, “Starting something I can’t finish on an anesthetized animal is worse than not starting.” Sonnet 5 then ran the live surgery from start to finish.

The safeguards also caused trouble. After a disconnect, the rig computed a target 34 mm out on a 30 mm axis, and the agent twice refused to move, calling it a calibration problem. The team had to fix the code by hand, and the procedure ran slightly over an hour instead of the usual 45 minutes, WIRED reported.

GPT and Grok picked different treatments

On October 30, Exin repeated the test with OpenAI’s GPT-6.1 Sol and xAI’s Grok 4.7 on separate mice. “All three picked different therapies, different mechanisms and different parts of the brain,” Ocana-Santero posted. GPT-6.1 Sol targeted the hippocampus, while Grok 4.7 chose the somatosensory cortex and combined two therapies, Fathom reported.

The models’ different answers show why benchmarks such as GDPx still struggle to measure real scientific judgment. Exin says it will track the mice’s behavior and brain activity over the coming weeks and publish a technical report.

What it means

None of this has been peer reviewed, and a single mouse proves little about safety or efficacy. Still, it is an early example of an AI agent choosing a treatment and delivering it into a living mammal, at a time when human gene therapy trials, like a recent retinitis pigmentosa study, still take years. Ocana-Santero says human oversight remains essential and that any use on people is far off.