Table of Contents
Gabriel Ocana Santero, founder and CEO of San Francisco-based biotech startup Exin Therapeutics, recently conducted an experiment where an artificial intelligence agent carried out a surgical brain injection on a live mouse. While the procedure required safety interventions and extra technical steps, it demonstrates how automated systems might eventually take a hands-on role in scientific laboratories.
Connecting AI Models Directly to Lab Instruments
Santero explained to WIRED that his interest stemmed from testing what happens when AI models interact directly with lab hardware rather than remaining limited to analyzing data or planning experiments. His startup uses AI to study brain activity to create treatments for neurodegenerative diseases and epilepsy. According to Santero, because AI is trained mainly on published literature, it misses real-time experimental nuances such as unexpected observations or failed trials. Connecting models directly to laboratory equipment could make them more reliable.
To test this concept, Santero’s team linked an AI agent to a surgical rig equipped with a camera, automated arms, a motorized immobilizer, and a needle injector. The test aimed to assess whether a gene therapy could lower seizure occurrences in an epileptic mouse. A veterinary surgeon performed the initial craniotomy, after which the AI model chose one of 19 potential treatments from a spreadsheet, selected the target brain coordinates, and operated the surgical instruments to deliver the substance.
Model Safety Refusals and Technical Adjustments
The team encountered unexpected hurdles with model behavioral guidelines. They initially attempted the procedure using Anthropic’s Claude Opus 5.5 model, but the software declined to guide the needle into the mouse’s tissue, citing risks to the animal. “He did the experimental planning, but when it was time to inject, he said, ‘something on my output is telling me I shouldn’t do this,’” Santero told WIRED, referencing the agent’s response.
The researchers then switched to Claude Sonnet 5, which agreed to proceed with the surgery. Anthropic did not respond to requests for comment regarding the models. To maintain safety, the team implemented precautions including mandatory human review of injection plans, restricted access to the hardware’s core software, and a physical emergency stop button to cut power to the motors.
During the surgery, a coordinate alignment error occurred. The AI agent detected the mistake, but its security boundaries prevented it from altering the core software to fix it. The researchers bypassed this by using a separate Opus 5 agent to modify the code without revealing that a live surgery was in progress. The troubleshooting extended the typical 45-minute procedure to slightly over an hour.
Future Tests and Human Oversight
Despite the delays, the research team was satisfied with the outcome and plans to evaluate other systems, including xAI’s Grok. They are also researching AI-controlled craniotomies, though live animal testing for that step is on hold until safety can be guaranteed. Santero emphasized that human supervision remains essential when working with live subjects, noting that the mouse involved in the trial is currently recovering well as researchers monitor the gene therapy’s effectiveness.
