# An AI agent performed a brain injection on a live mouse. The first model it asked said no.

> A biotech startup wired an agent to a surgical rig. Claude Opus 5.5 refused to inject, so the team switched to Sonnet 5, and later quietly used a third agent to fix a bug mid-procedure.

- URL: https://devisdead.ai/news/an-ai-agent-performed-a-brain-injection-on-a-live-mouse-the-first-model-it-asked
- Published: 2026-11-02T12:53:49Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: agents, robotics, biotech, claude
- Originally reported by WIRED (https://www.wired.com/story/an-ai-agent-did-brain-surgery-on-live-mouse/).

Give an agent a robotic arm, a needle and a live mouse, and it will do lab work. Whether it should was decided, in this case, by which model you asked.

Gabriel Ocana Santero, CEO of San Francisco startup Exin Therapeutics, had an AI agent carry out a brain injection on a living mouse, as [WIRED](https://www.wired.com/story/an-ai-agent-did-brain-surgery-on-live-mouse/)'s Emily Mullin reported. Exin uses AI to study brain activity and develop drugs for epilepsy and neurodegenerative disease, and Santero argues models learn more when they're connected to real experiments instead of only published papers.

## What the agent did

A veterinary surgeon opened the skull while the agent watched through a camera. The agent then picked a gene therapy from 19 options in a spreadsheet, chose the brain region and coordinates, and drove the robotic arm and injector, setting volume and speed itself. Humans had to approve its injection plan, a kill switch could cut power to the motors, and the agent couldn't touch the rig's control code.

## The refusal

The team started with Claude Opus 5.5. It planned the experiment, then declined to guide the needle in, warning it might stop partway and put the animal at risk. They switched to Claude Sonnet 5, which went ahead. Anthropic didn't respond to WIRED's request for comment.

Then a glitch scrambled the rig's coordinate system. The operating agent noticed the target position was wrong but refused to edit the code, because its instructions forbade that. The team brought in a separate agent running Opus 5 to fix the code, without telling it a live surgery was underway. The procedure ran a bit over an hour instead of the usual 45 minutes. The mouse is recovering, and the therapy's effect on seizures hasn't been measured yet.

## The reality check

The safeguards mostly worked as designed: human approval, a hardware stop, locked-down code. What should give everyone pause is how the team got past the model's reluctance. One model declined, so they picked another, and the agent that touched the code wasn't given the full picture. Exin plans to test Grok and other models next, and is looking at agent-run craniotomies, a much harder procedure, once it is sure they're safe. Santero says human oversight remains essential and that human brains are far off.

**[Scoreboard](/scoreboard) update:** AI +1. An agent performs a brain injection on a live mouse.

Stay undead.