An AI Agent Injected Drugs Into a Live Mouse’s Brain

AI agents can now perform complex surgical tasks on live animals, blurring the lines between AI and physical laboratory work.

Gabriel Ocana Santero, CEO of Exin Therapeutics, has demonstrated an AI agent's ability to perform a brain injection procedure on a live mouse. The experiment involved connecting an AI agent to a custom-built surgical rig equipped with a motorized device for immobilizing the mouse and a camera. The AI agent was tasked with selecting a gene therapy from a list, determining the injection site, and controlling robotic arms and an injector needle. Initially, Anthropic's Claude Opus 5.5 model refused to complete the injection, citing safety concerns. The team then switched to Claude Sonnet 5, which successfully executed the procedure after a veterinary surgeon performed a craniotomy. The AI controlled the injection parameters, but a technical glitch required a separate AI agent to fix the equipment's coordinate system, delaying the process. Safeguards included human approval for injection plans and an emergency stop button. Exin Therapeutics plans to test other AI models and is exploring agent-controlled craniotomies, emphasizing the need for human oversight in animal research.

This experiment signifies a significant step towards integrating AI agents as active participants in laboratory research, moving beyond design and analysis to direct physical intervention. By connecting AI to lab equipment, researchers aim to capture nuanced, real-time experimental data often missed in traditional AI training on published literature. This direct access could lead to more reliable AI models and accelerate drug discovery and development, particularly in complex fields like neuroscience.

The successful, albeit imperfect, execution of a brain injection by an AI agent suggests a future where AI collaborates more intimately with human scientists in physical tasks. Challenges such as model refusal due to safety protocols and technical glitches highlight the ongoing need for robust human oversight and sophisticated error correction mechanisms. The development of agent-controlled craniotomies points to increasingly complex autonomous laboratory procedures, though widespread deployment, especially in human research, remains distant.