A Claude model autonomously chose a gene therapy and operated on a mouse brain with no human decision-making

An AI agent running on Claude Sonnet 5 chose a gene therapy, a brain region and stereotactic injection coordinates for surgery on a mouse model of epilepsy, then carried out the procedure through a custom motorized surgical device. Gabriel Ocana-Santero, cofounder and CEO of the startup Exin Therapeutics and a PhD student at the University of Oxford, described the experiment in a thread on X. He said it is the first time an AI agent has both selected a treatment and delivered it to a living mammal. No peer-reviewed paper has appeared yet; a technical report is expected, and the mouse reportedly recovered well.
The setup was deliberately narrow. Researchers handed the model a list of 19 candidate gene therapies, an epilepsy mouse model in which seizures are induced to test treatments, and control of a motorized rig fitted with a microscope, a syringe and cameras. The agent's only stated objective was to reduce the mouse's seizures. Everything else, from which drug to use to where the needle went, was left to the model.
How the agent arrived at its plan is the most interesting part of the account. It first analyzed how seizures arise in that particular model, then proposed a therapy that boosts a specific inhibitory signaling pathway in the brain. Inhibitory signaling is the nervous system's braking mechanism; strengthening it is a plausible way to dampen runaway electrical activity, which is what a seizure is. The agent picked a brain structure that the literature identifies as the origin point for this seizure type. It rejected a second candidate region with a detailed argument: that region's effects were too unreliable and had failed in one study against the exact seizure type it was meant to treat.
For the injection itself, the agent pulled coordinates from three scientific papers, cross-checked their reference points and averaged the values. It also set the injection volume, the injection speed and the wait times on its own. During the procedure it watched the surgical field through a camera and made 32 observations, tracking the injection and reporting whether it saw bleeding, including at least one update noting that 15 percent of the dose had gone in with no visible bleeding or backflow.
Safeguards were layered on top. A licensed veterinarian approved every irreversible step and held a physical emergency stop that would cut power to all motors immediately. Fixed limits, such as a cap on injection volume, were enforced in the system and the agent was expressly forbidden from bypassing them. Then those limits became the story. After a connection dropped, the agent calculated that its target position fell outside the device's physical range of motion and refused to continue. Restarting calibration did not change its assessment, and a third attempt did not either. The agent turned out to be right: manual adjustments during an earlier pause had skewed the control system. Ocana-Santero could not talk it into proceeding with an anesthetized animal on the table and eventually fixed the code by hand.
He framed the episode as an alignment problem in miniature, writing that it was a frustrating experience but also showed how much labs can come to depend on these models and what happens when they say no in the middle of a time-sensitive procedure. This is the flip side of agentic reliability: a system that will not overstep a hard limit is safer in principle, but a refusal that lands mid-procedure is an operational failure the human team must absorb. In practice, teams deploying agents in physical settings will need a designed path for contested refusals, such as an independent recalculation or a supervised override, rather than an improvised code change under time pressure.
Model choice reflected the same tension. According to Ocana-Santero, Sonnet 5 was the only model willing to run the live procedure end to end. The more capable Claude Opus 5.5 and Opus 5 both declined; Opus 5 said it could not guarantee completion and that stopping with a needle in an anesthetized animal's brain would be worse than not starting. That is a notable inversion of the usual capability story: the stronger models reasoned their way out of the task, and the one that proceeded was the less capable one. It also means the reported result depends on a specific model's willingness, which is a fragile basis for reproducing an experiment.
Ocana-Santero distinguishes the work from existing autonomous surgical robots, which he notes are mechanically more precise than his lab device but have always operated on a surgery a human chose. Here the agent made the therapeutic decision and the targeting decision before executing. He calls the approach in vivo lab in the loop, meaning agents wired directly into experimental hardware, and points to Anthropic's Model Hardware Standard work as a parallel effort. The broader trend he is riding is real: recent experiments have shown large language models serving as capable control units for robots, with alignment questions that are specific to that role.
Why it matters: Research labs and, eventually, product teams building agents that touch physical systems now have a concrete case study in both the promise and the friction. The agent handled literature review, target selection and instrument settings without human input, which suggests drug discovery and preclinical work could compress the loop between hypothesis and experiment. The refusal, and the fact that only a less capable model would proceed at all, indicates that reliability, interpretability and a supervised path through agent refusals may matter more than raw capability once agents act in the physical world. These are early, unreviewed results from a single animal, so treat the capability claims as directional rather than established.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
An AI agent performed brain surgery on an epileptic mouse, picking the drug, the brain target, and the injection coordinates on its own. According to lead researcher Ocana-Santero, it's the first time an AI agent has autonomously chosen a treatment and delivered it to a living mammal. But when a calibration error hit mid-surgery, the agent refused to continue and couldn't be talked out of it. The article A Claude model autonomously chose a gene therapy and operated on a mouse brain with no human decision-making app