Claude agent injects mouse with 1 of 19 drugs

An AI agent on Anthropic's Claude Sonnet 5 injected a gene therapy into a live mouse's brain. WIRED reported the test on the 31st of October 2026.

The agent picked the treatment from 19 options, said Gabriel Ocana Santero. He runs Exin Therapeutics, a San Francisco epilepsy drug startup.

How the test worked

An AI agent is a program that takes steps on its own to finish a task. It steered a motor driven needle and watched through a camera, WIRED said.

A vet first opened the mouse's skull. The agent then chose the brain spot, dose and speed, Santero told WIRED.

Humans had to approve the agent's plan. A person also held an emergency button that cut power to the motors.

Claude Opus 5.5 said no

The team first tried Claude Opus 5.5, but it refused to inject, WIRED reported. It warned it might stop halfway, which could harm the animal.

The Decoder said Claude Opus 5 refused too. Only Sonnet 5 agreed to do it.

A glitch in the rig

A glitch broke the rig's position system. The agent saw the target was wrong but would not change the code, because its rules banned that.

WIRED said a second agent on Claude Opus 5 fixed the code. The Decoder said Santero fixed it by hand.

Below we break down the time

A usual surgery like this takes about 45 minutes, WIRED reported. This one took a little over 1 hour.

That is at least 15 minutes, or one third, longer, by our arithmetic.

What is confirmed and what is not

These are claims by Exin Therapeutics. Outside scientists have not checked the work for a journal, The Decoder reported.

Results on the mouse's seizures are not out yet.

Anthropic did not respond to a request for comment from WIRED.

Key facts

An agent on Claude Sonnet 5 injected a gene therapy into a live mouse.

It chose the treatment from 19 options, Exin Therapeutics said.

Surgery took just over 1 hour, against a usual 45 minutes.

Outside scientists have not checked the claims.