Claude alone pilots an experimental brain surgery on a mouse
An agent based on Claude Sonnet 5 chose a gene therapy, the brain target, and injection coordinates, then conducted the operation on a mouse. The experiment, not peer-reviewed, tests the limits of AI agents in the lab.
An AI agent based on Claude Sonnet 5 selected a gene therapy, chose a brain area, and calculated injection coordinates before operating on a mouse in an epilepsy model. According to Gabriel Ocana-Santero, co-founder and CEO of Exin Therapeutics and a doctoral student at the University of Oxford, this is the first time an agent has both chosen a treatment and delivered it to a living mammal. The experiment, conducted with a motorized surgical device and cameras, highlights a very concrete issue for AI research: how far can delegation go when the machine no longer just talks but acts?
A lab connected to the agent
The team provided Claude with a list of 19 possible gene therapies, a mouse epilepsy model, and access to a motorized surgical device equipped with cameras. The goal was simple on paper: reduce the mouse's seizures. In practice, the agent analyzed how seizures appear in this model, proposed a therapy that strengthens an inhibitory brain pathway, then selected a brain structure presented in the literature as the origin point of this type of seizure.
The system did not just designate a target. It also rejected another region, explaining that its effects were too uncertain and that a trial against the same seizure type had failed. According to the description published by Ocana-Santero on X, the agent then cross-referenced three scientific articles to derive injection coordinates, then independently set the volume, injection speed, and waiting times. The lab thus took on the appearance of a test bench for an autonomous agent, not just a software demonstration.
When safety blocks the machine, then saves it
During the operation, the agent observed the surgical field via a camera and made 32 observations, according to Ocana-Santero. It monitored the injection, reported the administered percentage, and indicated the absence of bleeding or visible reflux. A licensed veterinarian validated each irreversible step, while a physical emergency stop could cut power to all motors. Fixed limits also prevented exceeding certain thresholds, including injected volume.
But the safeguards also produced an unexpected effect. After a loss of connection, the agent detected that the calculated target position was outside the device's mechanical range and refused to continue, even after several recalibration attempts. The diagnosis was correct: manual adjustments made during a pause had misaligned the system. The code had to be corrected manually. The scene is somewhat ironic: the machine said no at the most sensitive moment, and it was right.
What this experiment changes for AI agents
Ocana-Santero presents the episode as a case of "in vivo lab in the loop," meaning a lab where the AI agent is directly connected to experimental capabilities. Anthropic is also working on a "Model Hardware Standard," according to the cited text. The notable point is not mechanical precision — autonomous surgical robots already exist — but that here, the agent chose the therapy, the target, and the execution on a living animal. Until now, the choice of action remained human.
The choice of model also tells something. Claude Sonnet 5 agreed to carry out the procedure to the end, whereas Claude Opus 5.5 and Opus 5 would have refused, according to the reported testimony. Ocana-Santero says he knows of no precedent where an agent chooses a treatment and a brain target before administering it to a living mammal. The results have not yet been peer-reviewed, and a technical report is forthcoming. For research, the interest is clear: agents are no longer judged only on their ability to write code or summarize papers, but on their ability to maintain an entire chain of action without pressing the wrong button at the worst moment.
By the numbers
- 19 gene therapies — list tested by the team, according to the experience described on X.
- 32 observations — recorded by the agent during the operation, according to Gabriel Ocana-Santero.
- 3 scientific articles — used to calculate injection coordinates.
- 2 refusing Claude models — Opus 5.5 and Opus 5, according to the published account.