AI Agent Successfully Injects Drugs into Live Mouse Brain
Executive Summary
"Exin Therapeutics has achieved a groundbreaking feat, using an AI agent to precisely inject drugs into a live mouse brain, heralding a new era of robotic surgery. This advancement blurs the lines between science fiction and reality, showcasing AI's potential in complex medical procedures."
Published: October 31, 2026
Imagine a future where robots, guided by artificial intelligence, perform delicate surgical procedures with precision far exceeding human capabilities. This isn't science fiction anymore. A San Francisco-based biotech startup, Exin Therapeutics, has taken a significant step toward this reality by demonstrating an AI agent's ability to successfully select and inject gene therapy into the brain of a live mouse. This groundbreaking achievement, while still in its nascent stages, offers a startling glimpse into how AI could fundamentally reshape laboratory research and, eventually, medical interventions.
src="https://media.wired.com/photos/6ae4d6bbd55a0e2bc0a8d990/master/w_1600,c_limit/103026-AI%2520Mouse%252002.jpg" alt="Photo-Illustration: Wired Staff; Getty Images"Photo-Illustration: Wired Staff; Getty Images
Gabriel Ocana Santero, CEO and cofounder of Exin Therapeutics, spearheaded the experiment. His motivation was clear: to explore the potential of AI agents when given direct access to laboratory equipment, moving beyond their traditional roles of experimental design and data analysis. "We’ve realized that we need to connect these agents to a lab to get the maximum value from them," Santero stated in an interview with WIRED.
AI Agent Takes the Scalpel (Figuratively)
Exin Therapeutics specializes in using AI to analyze brain activity for developing treatments for epilepsy and neurodegenerative diseases. Santero believes that current AI models, primarily trained on published literature, often miss crucial real-world research nuances – the failed experiments, unexpected observations, and the split-second decisions scientists make. By integrating AI with physical experiments, he aims to create more robust and reliable AI models.
To test this hypothesis, Santero's team devised a custom surgical setup. This rig included a motorized device to immobilize a live mouse, a camera for the AI to observe, and a robotic arm equipped with an injection needle. The AI agent's task was to perform a gene therapy injection into the brain of a mouse suffering from epilepsy, with the goal of reducing seizures.
src="https://media.wired.com/photos/6ae5110c5ad2ab653323a541/master/w_1600,c_limit/Surgical_rig_setup_mid_surgery.jpg" alt="The rig is controlled by an AI agent as it conducts brain surgery on a mouse."The rig is controlled by an AI agent as it conducts brain surgery on a mouse. Courtesy of Exin Therapeutics, Inc.
Navigating AI Hesitation and Glitches
The experiment wasn't a straightforward command-and-execute operation. The AI agent was tasked with selecting a treatment from 19 options presented in a spreadsheet. A veterinary surgeon first performed a craniotomy, creating an opening in the skull, after which the AI observed and then took control. It selected the therapy, target brain region, and precise injection coordinates. Crucially, it then controlled the robotic arms, the injector needle, and all injection parameters like volume and speed.
Interestingly, the team initially used Anthropic's Claude Opus 5.5. However, the AI refused to proceed with the injection, citing internal ethical safeguards, stating it shouldn't proceed midway through surgery due to potential harm 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 recounted. This hesitation, while problematic for the experiment, highlights the complex ethical considerations embedded in advanced AI systems.
The team then switched to Claude Sonnet 5, which was willing to proceed. While Opus is designed for heavy-reasoning and complex problem-solving, Sonnet is generally suited for more everyday tasks. This choice underscores the evolving capabilities and limitations of different AI models when faced with real-world, high-stakes scenarios.
The Road Ahead: Oversight and Future Potential
The procedure encountered further challenges. Despite built-in safeguards, including a requirement for human approval of the injection plan and an emergency stop button, a technical glitch disrupted the equipment's coordinate system. The AI correctly identified the error but was forbidden by its programming from directly modifying the faulty code. This issue was resolved by using a separate AI agent, unaware of the live surgery, to make the necessary code adjustments, causing a delay and extending the procedure beyond its typical 45-minute duration.
Despite these hurdles, Santero and his team were impressed by the AI's overall performance. They are now testing other models, including Grok, to identify the most effective AI for such tasks. Exin Therapeutics is also cautiously exploring agent-controlled craniotomies, a significantly more complex procedure, emphasizing that safety remains paramount before any live surgeries are attempted.
Santero envisions a future where AI agents actively collaborate with human scientists, directly operating lab equipment. For now, he stresses the indispensable role of human oversight, especially when dealing with living subjects. And, of course, the idea of unleashing an AI agent on a human brain remains firmly in the realm of distant future exploration.
As for the mouse, Santero reports it is recovering well, a testament to the careful planning and human intervention that guided this pioneering experiment.