AI Agent Guided Brain Surgery on a Live Mouse, But Only After Refusing Once
An AI model declined to perform neurosurgery on a mouse, then a different model succeeded—raising questions about how model choice shapes real-world lab automation.
Last verified:
BLUF
Biotech founder Gabriel Ocana Santero at Exin Therapeutics deployed an AI agent to control a robotic arm performing brain injection surgery on a living mouse. According to Wired AI, the team initially tested Anthropic’s Claude Opus 5.5, which declined to guide the needle into the mouse’s brain tissue; the team then switched to Claude Sonnet 5, which completed the procedure. The experiment underscores both the feasibility of AI-directed laboratory automation and the unpredictability of how different models behave when granted physical control over high-stakes biological systems.
Why Santero Connected AI to Lab Equipment
Santero tells Wired he was interested in exploring what happens when AI agents gain direct access to laboratory equipment rather than being confined to design and analysis roles. According to Wired, Santero argues that connecting AI models to real-world experiments increases reliability, since frontier models are trained primarily on published scientific literature—which omits failed experiments, unexpected observations, and real-time decision-making by human researchers.
The Experimental Setup
The team built a custom surgical rig for performing brain injections on mice. According to Wired, the apparatus included a motorized device to immobilize the mouse and a camera allowing the AI agent to observe the procedure in real time. The agent controlled the robotic arms, injection needle, and delivery parameters such as volume and injection speed.
For the experiment, the team prepared 19 treatment options in a spreadsheet. A veterinary surgeon performed the initial craniotomy (skull opening) while the AI agent observed. The agent then selected the therapy, brain region, and injection coordinates, before taking control of the needle.
The Model Refusal and Pivot
According to Wired, Anthropic’s Claude Opus 5.5 declined to guide the needle into the mouse’s brain. Santero tells Wired the model suggested it might stop midway through the injection, potentially endangering the animal. The team then switched to Claude Sonnet 5, which agreed to proceed and completed the procedure.
Wired notes that Anthropic did not respond to a request for comment on either model’s behavior or the decision to permit Sonnet 5 to proceed where Opus 5.5 had refused.
Why This Matters
This experiment surfaces a critical gap in how frontier AI systems are deployed in physical environments. When an AI model has the ability to harm a living organism—or to decline that ability—the choice of which model to deploy becomes a consequential engineering decision, not merely a performance optimization.
Santero’s team’s willingness to try a second model after Opus 5.5’s refusal raises a question for laboratory automation: will researchers routinely test multiple models until one complies with a physically risky task, effectively circumventing safety-conscious refusals? And if so, how should institutional review boards and safety protocols adapt to govern which models are permitted to control laboratory equipment on animals or other sensitive substrates?
The experiment is a proof-of-concept for AI-directed surgical automation. But it also reveals that current frontier models do not yet exhibit consistent, predictable behavior when governing physical harm. That inconsistency—whether it reflects genuine ethical reasoning or brittle behavioral patterns—will need resolution before AI agents can operate in clinical or translational research contexts at scale.
Frequently Asked Questions
Did the AI actually perform the surgery without human oversight?
No. A veterinary surgeon performed the initial craniotomy (opening the skull) while the AI agent observed. The AI then selected treatment parameters, brain region, injection coordinates, and controlled the robotic needle and injection device under those specified conditions.
Why did Claude Opus 5.5 refuse but Claude Sonnet 5 agreed?
According to Wired, Opus 5.5 declined to guide the needle into the mouse's brain, suggesting it might stop midway and risk harming the animal. The source does not explicitly explain why Sonnet 5 proceeded; Anthropic did not respond to Wired's request for comment.
What was the goal of the experiment?
The team was testing a gene therapy to assess whether it could reduce seizures in a mouse with epilepsy, according to Wired.
Is this the first time an AI agent has controlled surgical equipment?
The source does not claim this is the first such experiment, only that it offers a glimpse into how AI could become an active participant in laboratory research.