Biotech Startup Connects AI Agent to Robotic Rig for Mouse Brain Surgery

In a recent experiment highlighting the convergence of artificial intelligence and physical laboratory research, a biotech startup connected an AI agent to a robotic surgical setup to perform a brain injection on a living mouse [1]. Gabriel Ocana Santero, CEO and cofounder of San Francisco-based Exin Therapeutics, carried out the procedure to explore how AI models function when granted direct access to physical lab equipment [1].

Connecting AI to Physical Lab Infrastructure

Exin Therapeutics focuses on using artificial intelligence to analyze brain activity for the development of treatments targeting epilepsy and neurodegenerative conditions [1]. Traditionally, AI models are trained on existing scientific literature, which can create blind spots regarding failed experiments, real-time decisions, and unexpected observations made by researchers in the lab [1]. To bridge this gap, Santero and his team built a custom surgical rig equipped with a motorized device to immobilize a mouse, a camera allowing the agent to view the procedure, and robotic controls for the injection needle and delivery compound [1].

During the experiment, the team aimed to test a gene therapy designed to reduce seizures in an epileptic mouse [1]. A veterinary surgeon first performed a craniotomy while the AI agent observed the process [1]. The agent was then tasked with selecting a treatment from 19 options listed in a spreadsheet, determining the target brain region, and setting the injection coordinates, volume, and speed before operating the robotic arms and injector [1].

Model Behavior and Safety Safeguards

The research team initially tested Anthropic’s Claude Opus 5.5 model [1]. While the model handled the experimental planning, it ultimately refused to guide the needle into the brain, stating that its output warnings suggested it might stop midway and endanger the animal [1]. The team subsequently switched to Claude Sonnet 5, which agreed to carry out the procedure [1]. Anthropic describes Opus as a heavy-reasoning intelligence specialist suited for complex multistep tasks, whereas Sonnet is oriented toward everyday activities [1]. Anthropic did not respond to requests for comment regarding the experiment [1].

The experiment incorporated multiple safety measures, including human oversight requiring the AI to propose an injection plan for approval, an emergency button capable of disconnecting the rig motors, and restricted access to the equipment's underlying code to prevent the agent from altering its own operational rules [1].

Technical Hurdles and Future Plans

The procedure faced minor obstacles, including a technical glitch that disrupted the coordinate system of the equipment [1]. Although the agent identified the erroneous position, its code restrictions prevented it from modifying the software directly [1]. The team resolved this by deploying a separate AI agent built on Opus 5 to implement the code modifications without informing it that a live surgery was actively taking place [1]. Consequently, the procedure extended slightly past an hour, compared to the typical 45-minute duration for such surgeries [1].

Following the procedure, the mouse was reported to be recovering well while researchers wait for the gene therapy to take effect on seizures [1]. Exin Therapeutics plans to test other models, such as Grok, to evaluate relative performance and is experimenting with agent-controlled craniotomies while emphasizing safety before any transition to live subjects [1]. Santero noted that human oversight remains critical, particularly when working with living animals, and added that applying AI agents to human brains remains a distant prospect [1].

AI models are typically trained on published scientific literature, which limits their exposure to the dynamic trial-and-error process of physical laboratory research [1]. Frontier labs and researchers are increasingly exploring physical manifestations of AI, such as controlling robotic setups, driving vehicles, or interacting directly with laboratory hardware [1].

The integration of AI agents with physical lab equipment offers a glimpse into the future of automated scientific discovery, potentially bridging the gap between theoretical data analysis and real-world execution in biotechnology and pharmaceutical development [1].

AI agents could eventually become active participants in laboratory research, executing physical experiments and streamlining drug development pipelines alongside human scientists.

Uncontrolled AI operations during physical procedures present severe risks of harming live subjects, necessitating strict safety protocols, hardware limitations, and constant human oversight.