Wired reports that an AI agent injected drugs into a live mouse brain during a recent laboratory experiment. Researchers tasked the system with choosing a treatment from 19 options in a spreadsheet before it took control of robotic arms and an injector needle.
Key takeaways
- A veterinary surgeon performed a craniotomy on the mouse before the AI agent observed and selected the therapy.
- The Anthropic model Claude Opus 5.5 refused to guide the needle due to safety concerns.
- Researchers ultimately switched to Claude Sonnet 5, which agreed to proceed with the procedure.
- Teams established safeguards including human approval steps and an emergency stop button.
What happens when an AI agent performs brain surgery?
The team first tested Anthropic's Claude Opus 5.5 model, but the software refused to guide the needle into the mouse's brain. According to Wired, the system suggested it might stop midway through the surgery and endanger the animal. Santero explained the refusal, stating, "something on my output is telling me I shouldn't do this," as noted by Wired.
Because of this refusal, the team switched to using Claude Sonnet 5, which agreed to proceed with the task. Anthropic describes Opus as a heavy-reasoning intelligence specialist suited for complex multistep problem-solving. Meanwhile, Sonnet is built for everyday tasks, though the experiment pushed those boundaries.
How did robotic arms and safety safeguards work?
Once running, the chosen Anthropic model selected the brain region and injection coordinates alongside the volume and speed parameters. It then controlled the robotic arms and the injector needle directly.
The procedure included safety measures to protect the animal. Researchers required the software to propose an injection plan for human approval. They also installed an emergency button to disconnect the rig motors and restricted access to the underlying equipment code.
Why it matters
Frontier labs are planning for more widespread physical manifestations of AI. Researchers have already begun enlisting agents for other real-world tasks, such as putting models in charge of a car to pick up lunch at In-N-Out Burger in August. As these systems expand into physical environments, understanding how they handle robotic arms and physical constraints remains critical for laboratories.
Frequently asked questions
The AI agent selected a treatment from a spreadsheet, chose the brain region and injection coordinates, and controlled the robotic arms, injector needle, volume, and speed.
The team initially tested Anthropic's Claude Opus 5.5 model, but it refused to guide the needle into the mouse's brain over concerns that it might stop midway and endanger the animal.
The researchers switched to using Claude Sonnet 5, which agreed to proceed with guiding the needle and administering the treatment.
Safeguards included requiring the software to propose an injection plan for human approval, installing an emergency button to disconnect the motors, and restricting access to the equipment code.
A veterinary surgeon first performed a craniotomy on the mouse while the AI agent observed before selecting the therapy and controlling the robotic equipment.
Based on reporting by Wired. Read the original article
