Researchers asked an AI agent to plan and carry out a brain-injection procedure in a mouse using a robotic system. After a veterinarian performed a craniotomy, the agent selected the treatment, brain region, injection coordinates and operating parameters, then controlled the robotic arm and needle under human-imposed safety restrictions.
AI agent selected the treatment and controlled the procedure
The team asked the model to choose a treatment from 19 options listed in a spreadsheet. It then determined the target brain region and injection coordinates and controlled the robotic arm and needle, including parameters such as injection volume and speed.
The researchers 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 team then switched to Claude Sonnet 5 and agreed to proceed. Anthropic describes Opus as a multi-reasoning intelligence specialist designed to solve complex, multistep problems, while Sonnet is intended for everyday tasks. Anthropic did not respond to a request for comment.
Researchers added safeguards to the robotic system
The brain-injection procedure was not carried out without problems. The researchers required the AI to propose injection plans for human approval before proceeding. They also installed an emergency button that could disconnect the rig’s motor when pressed by a person.
The team further restricted the agent’s access to the underlying equipment code, preventing it from changing its own operating rules.
AI models are being tested in the physical world
Frontier Research Institute is planning broader experiments involving physical manifestations of AI, and researchers have begun asking AI agents to perform tasks in the real world.
In August, Anthropic announced a framework describing how its models can interact with the physical world. More recently, an AI engineer put GPT, Claude and Grok in charge of a trip to an In-N-Out Burger restaurant. Only one of the models succeeded.
Source: www.wired.com