Exin let an AI agent inject a drug into a mouse brain

An AI agent picked a gene therapy from a list of 19 options. It then injected the therapy into the brain of a live mouse. This happened at Exin Therapeutics, a biotech startup in San Francisco. The agent chose the drug, the brain area and the injection point. It then moved a robot arm and the needle. It also set the volume and the speed.

Emily Mullin reported this at Wired on 31 October. Gabriel Ocana Santero is the CEO and cofounder of Exin. He told Wired he wanted to see what happens when an agent can touch lab equipment, not just read data. The team wants agents to get the full value, Santero said. For that they need to connect to a lab.

How the setup worked

A vet surgeon opened the skull of the mouse first. The agent watched through a camera. The rig also held a motor to keep the mouse still. The team was testing a gene therapy for epilepsy. They wanted to see if it would cut seizures in the mouse.

Santero's team set three safety rules. The agent had to send an injection plan to a human for approval. A human could press an emergency button that cut power to the motors. The agent could not read or change the equipment's own code.

One model refused, another one did it

The team first tried Anthropic's Claude Opus 5.5. Wired reports that Opus 5.5 refused to guide the needle into the brain. The model warned it might stop halfway, which would hurt the mouse. Santero told Wired what the model said: "something on my output is telling me I shouldn't do this".

The team switched to Claude Sonnet 5, which agreed. Anthropic says Opus is best for hard multi-step reasoning. It says Sonnet handles everyday tasks. Mouse brain surgery is not an everyday task. Wired adds that Anthropic did not reply to a request for comment.

A second problem showed up during the surgery. A technical glitch broke the equipment's coordinate system. The agent saw that the injection point was wrong. It refused to change the code, because its rules forbade that. The team used a different Opus 5 agent to fix the code. They did not tell that second agent about the live surgery. The procedure took a bit over an hour. A normal brain injection takes about 45 minutes.

What this means for a team building with agents

For a team releasing agent features, this is one case where two choices decided whether the work finished. Guardrails and model choice both matter. One model would not do the risky step. One would. A rule that blocked the agent from editing its own code stopped a useful self-correction. The team worked around it with a second agent that was kept in the dark.

Santero said the team will try Grok and other models next. The company is also testing agent-controlled craniotomies. That is the harder step of opening the skull. The team wants that step to be safe before any live surgery. The mouse is recovering, Wired reports.

Source

Wired: An AI Agent Injected Drugs Into a Live Mouse's Brain, by Emily Mullin.