Source time Edition date
IndustryWired AI
An AI Agent Injected Drugs Into a Live Mouse’s Brain
AI智能体向活体小鼠大脑注射药物
Wired AI reports a biotech startup letting an agent choose a gene therapy and inject it into a live mouse, a tradeoff between faster lab iteration and who signs off before an automated choice reaches a living animal.
Judgment
Gabriel Ocana Santero, CEO of Exin Therapeutics, must decide whether to extend the agent from injection to agent-controlled craniotomies and live surgeries, accepting added delay and mandatory human approval plus an emergency stop as the price of faster iteration.
Basis
Santero told WIRED that Exin connected an AI agent to a custom rig—motorized immobilizer, camera, needle, injector—because models trained on published literature miss failed experiments and real-time decisions. The team asked the agent to choose a treatment from 19 spreadsheet options; after a veterinary surgeon performed the craniotomy, the agent selected the therapy, brain region, injection coordinates, volume, and speed. Claude Opus 5.5 refused to guide the needle, so they switched to Claude Sonnet 5, which proceeded. Safeguards required human approval of the injection plan, an emergency button to disconnect the motors, and restricted access to equipment code. A coordinate-system glitch still delayed the procedure: a typical brain injection takes about 45 minutes, while this run lasted slightly longer than an hour; a separate Opus 5 agent fixed code without being told live surgery was happening. Santero says they are waiting for the therapy to take effect and testing Grok next.
Unclear The source does not say whether the gene therapy reduced seizures, what dose or vector was injected, or whether an animal-care review board approved the protocol.
Next Watch for Exin's seizure results from this mouse and for a published protocol covering human approval and emergency-stop rules before any agent-controlled craniotomy on a live animal.