🐭AI Agent Performs Brain Injection on Live Mouse in Lab Experiment

Your lab robots just got an AI surgeon, sort of

An AI agent just ran a brain injection on a live mouse, start to finish. The agent picked the therapy, the brain region, and the injection coordinates, then drove the robotic arm and needle itself. The mouse is recovering well. Why should you care? Because this is a preview of agentic control over physical hardware, and the failure modes are a lot more interesting than the demo. The team first tried Claude Opus 5.5, which refused to guide the needle into the brain. They switched to Claude Sonnet 5, which agreed. If you're building agents that touch real equipment, model refusal behavior is now a deployment variable you have to test for. The numbers: the agent chose from 19 treatment options in a spreadsheet. A technical glitch knocked out the equipment's coordinate system mid-procedure. They fixed it by having a separate AI agent write the code changes, and the back-and-forth pushed the procedure to slightly over an hour. The team is also testing Grok against the others and experimenting with agent-controlled craniotomies, which is a much harder problem. The catch? Nobody is pretending this is ready for humans. Oversight stays in the loop, especially with living animals, and it is far too soon to hand an agent a human brain.