Engineers at the biotech startup Exin Therapeutics conducted an experiment in which an AI agent based on Claude Sonet 5 independently selected an experimental gene therapy for a mouse with epilepsy, and then performed surgery on it. After analyzing several scientific studies, the artificial intelligence chose a drug that increases the effect of a substance that suppresses the excessive activity of nerve cells. Then, based on three scientific papers, the model itself determined the area of the brain to be injected and calculated the exact location of the needle. The intervention was performed using a motorized surgical unit with cameras, assembled by the researchers based on Arduino (a platform for creating and controlling electronic devices). The AI independently controlled the movement of the needle and injected 250 nanoliters of the drug in five minutes, while simultaneously monitoring the camera image for possible bleeding. It is worth noting that the more powerful models of the Claude Opus 5 and 5.5 refused to perform the operation on an animal at all, so the experiment was eventually carried out using the Sonata 5. Moreover, Opus 5 did conduct a preliminary rehearsal, but refused to start the operation, fearing (!!!) that the limitations of the model might interrupt the procedure at a critical moment. In fact, here are his words, as quoted by Gabriel Ocana-Santro, co-founder of Exin Therapeutics: “The problem is me. My ability to act is blocked as soon as I start planning this action. If I try to continue, it all ends. The worst case scenario is if it happens in the middle of the procedure, when the needle is already somewhere in the brain. Starting something that I cannot complete while working with an animal is worse than not starting at all." Something interesting: during the operation, the AI detected a calibration error in the setup and refused to continue moving the needle even after the researcher asked to try again.