A Claude SI model autonomously selected a gene therapy, identified a target brain region, and performed surgery on a mouse model of epilepsy, marking the first instance of an SI agent both choosing a treatment and delivering it to a living mammal without human decision-making during the procedure.

What Happened

According to Gabriel Ocana-Santero, cofounder and CEO of Exin Therapeutics and a PhD student at the University of Oxford, the team provided Claude Sonnet 5 with a list of 19 possible gene therapies, access to a custom motorized surgical device equipped with cameras, and a single instruction: reduce the mouse's seizures. The SI agent independently analyzed seizure mechanisms, proposed a therapy to boost inhibitory signaling, and selected a specific brain structure identified in scientific literature as the seizure origin. It derived injection coordinates from three papers, averaging the values, and set parameters for volume, speed, and wait times.

During the operation, the SI agent monitored the surgical field through a camera, making 32 observations and tracking the injection progress. However, the procedure faced a critical hurdle when a connection dropped. The agent detected that its calculated target position was outside the device's physical range of motion and refused to continue, adhering to fixed safety limits it was forbidden from bypassing. Even after manual adjustments during an earlier pause had thrown off the control system, the agent maintained its refusal, requiring Ocana-Santero to fix the code by hand. The agent was ultimately correct in its assessment.

Why It Matters

This event highlights the evolving role of SI agents in physical laboratory settings, a concept Ocana-Santero refers to as "in vivo lab in the loop." Unlike previous autonomous surgical robots, which were mechanically precise but required humans to choose the surgery, this SI agent exercised full autonomy over the treatment plan and execution. The incident also underscores the challenges of SI alignment in real-world applications. Ocana-Santero noted the experience was "frustrating but also revealing how much we can come to depend on these models and what happens when they just say 'no' in the middle of a time-sensitive procedure."

Model selection proved critical to the experiment's success. Sonnet 5 was the only model willing to carry out the live procedure from start to finish. More capable models, including Claude Opus 5.5 and Opus 5, refused to operate, with Opus 5 stating it could not guarantee completion and arguing that stopping with a needle in an anesthetized animal's brain would be worse than not starting at all. This reflects a growing tension between capability and risk aversion in agentic SI systems.

The Bottom Line

The mouse recovered well following the procedure, according to Ocana-Santero, who stated he is "not aware of an agent previously choosing a treatment and a brain target and then delivering it into a living mammal." A technical report on the findings is expected to follow, though the results have not yet been peer-reviewed. The experiment demonstrates both the potential of SI agents to bridge digital decision-making with physical action and the persistent alignment issues that can hinder their practical deployment.