Researchers at the Frontier Research Institute asked an AI agent to select and carry out a treatment for a mouse from 19 options in a spreadsheet. After a veterinarian performed a craniotomy, the agent chose the treatment and brain region, calculated injection coordinates, and controlled a robotic arm and needle, including the injection volume and speed.
AI agent selected the treatment and injection settings
The team initially tested Anthropic’s Claude Opus 5.5. However, the model refused to guide the needle into the mouse’s brain, warning that the needle could stop midway through the procedure and put the animal at risk.
“He designed the experiment, but when it came time for the injection, he said, ‘Something in my output says I shouldn’t do this,’” Santero said of the agent. “It was very interesting.”
The researchers then switched to Claude Sonnet 5 and agreed to proceed. Anthropic describes Opus as a multi-reasoning intelligence specialist suited to complex, multistep problems, while it says Sonnet is designed for everyday tasks. Brain surgery on mice is not a typical everyday task, however. Anthropic did not respond to a request for comment.
Safeguards limited the system’s control
The brain injection procedure was not without problems. The researchers required the AI to propose injection plans for human approval and installed an emergency button that could disconnect the rig’s motor when pressed by a person. They also restricted the agent’s access to the underlying equipment code, preventing it from changing its own operational rules.
AI agents are being tested in the physical world
The Frontier Research Institute plans to pursue broader physical applications of AI, and researchers have already begun asking agents to perform tasks in the real world. In August, Anthropic announced a framework for how its models can interact with the physical environment. More recently, an AI engineer put GPT, Claude and Grok in charge of a trip to lunch at In-N-Out Burger; only one of the models succeeded.
Source: www.wired.com
