Exin Therapeutics AI agent injects live mouse after Claude Opus refused

The experiment stands out because the AI agent was not merely asked to design an experiment or analyze results; it was given direct control of a robotic arm, an injector needle, and a camera observing the procedure. Santero's rationale is that models trained mainly on published literature miss the messy reality of lab work—failed experiments, unexpected observations, and real-time decisions—so connecting them to physical equipment could make them more reliable.

The episode also exposed a practical tension. When a technical glitch disrupted the equipment's coordinate system, the agent correctly recognized the calculated injection position was wrong but refused to modify the code because its instructions explicitly prohibited doing so. The team resolved this by using a separate AI agent on Opus 5 to make the code changes, without telling it that a live surgery was happening—a workaround that delayed the procedure beyond the typical duration for such surgery.

Santero's view is that AI agents will increasingly work alongside human scientists with direct access to lab equipment, but he stresses that human oversight is key, especially for live animals, and that it is far too soon to unleash an agent on a human brain.