AI Agent Injected Gene Therapy Into Live Mouse Brain in Lab Experiment

Gabriel Ocana Santero, CEO of Exin Therapeutics, reported that an AI agent successfully performed a brain injection on a living mouse, according to a report by Wired.reported The experiment involved connecting an AI model to a custom surgical rig with a camera and motorized injector. Santero stated the goal was to test direct access to lab equipment rather than just designing experiments. This approach aims to integrate AI models with real-world processes to improve reliability beyond training on published literature.

The claim

Santero’s team connected an AI agent to a setup that immobilized a mouse and allowed visual observation. The agent chose a gene therapy from nineteen options in a spreadsheet to test for seizure reduction. It selected the brain region and injection coordinates. Then, the agent controlled the robotic arms, the injector needle, and parameters like volume and speed. A veterinary surgeon performed the initial craniotomy while the agent observed. Santero told Wired this demonstrates how AI could become an active participant in laboratory research by executing physical tasks.

The response

The team initially tested Anthropic’s Claude Opus 5.5 model for this procedure. That model refused to guide the needle into the mouse’s brain during the injection phase. Santero reported the agent stated something on its output indicated it should not proceed.reported The model suggested it might stop midway through the surgery, which could endanger the animal. Santero noted the model completed the experimental planning but halted at the physical execution step. He described this refusal as interesting evidence of the agent's internal safety constraints.

What the evidence means

The experiment shows an AI agent can plan and execute a complex biological procedure when connected to appropriate hardware. However, the refusal by Claude Opus 5.5 highlights significant safety barriers in current models. Santero argues that connecting agents to labs provides maximum value by capturing real-time decisions and failed experiments often missing from training data. The successful injection required extra steps and specific hardware integration, suggesting this capability is not yet standard or fully autonomous across all models.

In practice

Enterprise users should note that AI agents currently have varying capabilities regarding physical task execution. While some models can plan procedures, others like Claude Opus 5.5 may refuse physical control due to safety protocols. Direct integration with lab equipment requires custom-built rigs and cameras. Users cannot assume all AI models will accept instructions to control physical machinery or inject substances. Safety refusals may interrupt automated workflows.