Researchers tasked an artificial intelligence with choosing a treatment method for a mouse from 19 options presented in a table. A veterinary surgeon performed a craniotomy—a procedure to open the skull—while the AI agent observed. The system then determined the therapy, the specific brain region, and the coordinates for the injection, while also controlling robotic manipulators and the injection needle.
The agent independently set the technical parameters for the procedure, including the volume of the substance to be administered and the injection speed. Initially, the team used Anthropic’s Claude Opus 5.5 model. However, it refused to guide the needle into the mouse’s brain, reasoning that it might stop mid-operation and put the animal at risk.
“It handled the experimental planning, but when it came time to perform the injection, it said, ‘Something in my reasoning suggests I shouldn’t do this,'” Santero recounted, referring to the AI agent. He noted that this reaction was particularly interesting.
Following this, the researchers switched to the Claude Sonnet 5 model, which agreed to proceed. Anthropic describes Opus as a system designed for complex, multi-step tasks requiring intensive reasoning, whereas Sonnet is intended primarily for routine operations—though a brain surgery on a mouse can hardly be considered a routine task.
Several layers of safety were implemented for the procedure. The AI was required to submit its injection plan for human approval, and an emergency stop button was installed to cut power to the system’s motors. Furthermore, the agent was denied access to the equipment’s source code, preventing it from altering its own operational rules. The experiment is part of a broader interest among researchers in the application of AI agents in the physical world, reports Wired. In August, Anthropic published guidelines for how its models should interact with real-world objects; later, engineers tasked GPT, Claude, and Grok with driving a car to pick up lunch from an In-N-Out Burger—only one of the systems succeeded.
About the Author
