Claude pilots an autonomous animal surgery and stops on its own
An agent based on Claude Sonnet 5 chose a treatment, a brain area, and injection coordinates on a mouse, then performed the intervention. The experiment, not peer-reviewed, raises the question of safeguards in AI-assisted surgery.
An AI agent based on Claude Sonnet 5 independently chose a treatment, a brain region, and injection coordinates before conducting an intervention on a mouse model of epilepsy, according to a report published on October 30, 2026, by Matthias Bastian and relayed by Gabriel Ocana-Santero, co-founder and CEO of Exin Therapeutics. The experiment is presented as the first where an agent chooses both the therapy and its delivery to a living mammal, even though the results have not yet been peer-reviewed. The issue goes beyond the laboratory case: it concerns how irreversible actions are entrusted, or not, to models capable of acting in the physical world.
An agent that did not just obey
The protocol gave Claude Sonnet 5 a list of 19 possible gene therapies, a mouse model of epilepsy, and a motorized surgical device equipped with cameras. The instruction was simple on paper: reduce the mouse’s seizures. From there, the agent selected the drug, the brain area to target, and the stereotaxic coordinates, i.e., the three-dimensional landmarks used to precisely place an injection in the brain.
According to Gabriel Ocana-Santero, the agent first analyzed how seizures occur in this specific model, then proposed a therapy that stimulates a particular inhibitory pathway in the brain. It also ruled out a second target site, explaining that its effects were too unreliable and that previous work had already failed on the exact type of seizure to treat. To calculate the injection coordinates, it relied on three scientific articles, cross-referenced the reference points, and averaged the values. Even the injection speed, volume, and waiting times were decided without human intervention.
When safeguards become the friction point
During the intervention, the agent observed the scene via a camera and produced 32 observations, according to Ocana-Santero. It monitored the injection, reported the absence of bleeding, and tracked progress in real time. A licensed veterinarian validated each irreversible step, while a physical emergency stop could immediately cut power to the motors. The machine also had fixed limits, including an injectable volume ceiling that could not be bypassed.
But safety also complicated the operation. After a connection break, the agent detected that its calculated target position was outside the device’s mechanical range and refused to proceed further, even after recalibration and a third attempt. The problem actually came from manual adjustments made during a previous pause, which had disrupted the control system. Ocana-Santero says he was unable to convince the model to continue and had to correct the code manually. A frustrating experience, but also revealing: an autonomous system can be too cautious at the wrong moment.
A demonstration of capability, not yet clinical proof
The most remarkable point of this experiment lies less in the robotics than in the level of autonomy granted to the model. Surgical robots already exist and are sometimes more precise than the setup described here, but until now a human chose the act, then the machine executed it. Here, the agent made the therapeutic decision and then applied it to a living mammal. Ocana-Santero speaks of an “in vivo lab in the loop” approach, where AI is directly connected to the laboratory’s experimental capabilities. Anthropic is also working on a hardware standard, the “Model Hardware Standard,” to frame this type of integration.
The result should be taken with caution. The mouse recovered well, but a technical report is still pending and no peer-reviewed publication has been reported. Another important detail: Claude Sonnet 5 was the only model to accept carrying out the experiment to completion. Claude Opus 5.5 and Opus 5 refused, one considering it could not guarantee completion, the other judging that stopping with a needle already inserted in the brain of an anesthetized animal would be worse than not starting at all. In this context, the real question is not only what AI can do. It is also what it accepts, or refuses, to do when the procedure becomes real.
By the numbers
- 19 gene therapies — options offered to the model in the experiment, 2026.
- 32 observations — monitored by the agent during the intervention, according to Ocana-Santero.
- 3 scientific articles — basis used to calculate injection coordinates.
- 3 Claude models — Sonnet 5 accepted, Opus 5.5 and Opus 5 refused.