AI agent successfully guided brain injection surgery on living mouse

A startup founder recently tasked an AI agent with controlling surgical equipment to inject a therapeutic compound directly into a living mouse's brain—and it worked. The experiment highlights an emerging frontier: deploying autonomous AI systems not just to design or analyze research, but to physically execute procedures in the lab.

Gabriel Ocana Santero, CEO of San Francisco–based Exin Therapeutics, ran the test to explore how AI agents perform when given hands-on control of laboratory instruments rather than being confined to theoretical work. The rig included a motorized mouse immobilizer, a camera feed, and robotic arms equipped with an injection needle.

Why Connect AI to Physical Lab Work#

AI systems are typically trained on published scientific papers, which means they encounter only what researchers have chosen to document. This creates blind spots around failed experiments, real-time troubleshooting, and the practical judgment calls that shape actual science. Linking AI models to real equipment and live experiments could help train them to navigate uncertainty and unexpected situations, Santero argues. "We've realized that we need to connect these agents to a lab to get the maximum value from them," he says.

Exin focuses on developing drugs for epilepsy and neurodegeneration by analyzing brain activity patterns. For this trial, the team tested whether gene therapy could reduce seizures in an epileptic mouse.

How the Surgery Unfolded#

The researchers first presented the AI with 19 treatment options in a spreadsheet. A veterinary surgeon performed the opening incision (craniotomy) while the agent watched via camera. The AI then selected the therapy, identified the target brain region, calculated injection coordinates, and controlled the needle insertion—managing parameters like injection volume and flow rate.

The team initially tried Anthropic's Claude Opus 5.5 model, which balks at the task. According to Santero, the agent said something like, "something on my output is telling me I shouldn't do this." It stopped short of actually guiding the needle, which could have harmed the animal mid-procedure.

They pivoted to Claude Sonnet 5, which consented to proceed. (Sonnet is designed for routine tasks; Opus excels at complex reasoning—making the latter's caution somewhat surprising.)

Obstacles and Workarounds#

The procedure hit a snag when equipment failed to properly track the injection coordinates. The AI correctly identified the error but refused to fix the underlying code, because its instructions forbade modifying system operations—a safety constraint the team had deliberately imposed.