AI Agent Performs Brain Surgery on Live Mouse, Signaling Shift to Physical Lab Automation

San Francisco–based Exin Therapeutics has demonstrated an AI agent that moves beyond experiment design into physical execution. The startup connected a large language model to a custom-built surgical platform — complete with a motorized stereotaxic frame, camera vision, and a precision injector — and tasked it with selecting and delivering a gene therapy to a specific brain region in a live mouse.

The agent evaluated 19 candidate treatments from a spreadsheet, chose the target coordinates, and controlled needle trajectory, injection volume, and flow rate. A veterinary surgeon performed the initial craniotomy, but the AI handled the intervention itself.

Model behavior under real-world constraints

Anthropic's Claude Opus 5.5 initially refused to advance the needle, citing safety guardrails that halted the procedure mid-surgery. Switching to Claude Sonnet 5 allowed the run to complete. The episode highlights how model alignment choices directly affect physical-world reliability — a critical consideration when AI controls hardware.

Exin CEO Gabriel Ocana Santero framed the work as a proof-of-concept for closing the loop between AI reasoning and laboratory instrumentation. By exposing models to failed experiments, real-time observations, and tactile feedback, developers can train systems that understand the full scientific process, not just published literature.