Exin Therapeutics Uses AI Agent to Perform Brain Surgery on a Mouse; Opus 5.5 Refuses to Execute

Claude performed brain surgery on a live mouse, but Opus 5.5 refused to do it APPSO October 31, 2026 10:09:20

At a startup in San Francisco called Exin Therapeutics, on a motorized surgical platform built from a modified Arduino, an AI Agent performed brain surgery on a mouse.

So far, the mouse is recovering well.

An AI Agent performed brain surgery on a mouse

The research team prepared a spreadsheet for the Agent listing 19 gene therapy vectors; a mouse known to have seizures, a surgical platform equipped with a camera, and all that was left was a single task instruction: reduce seizures.

The surgical platform was a homemade device the team built with Arduino, and the linkage between the camera and the motors was also wired in with their own custom setup. In their description, the team joked that this was a makeshift rig dedicated to "lab closed-loop" enthusiasts, and that they were still waiting for an early version of Anthropic's official MHS system, hoping to switch to more professional tools soon.

Next, the Agent first analyzed how seizures spread in this model, chose a vector that could enhance inhibitory signals in GABAergic neurons, and set the injection site at the lateral entorhinal cortex, because the literature said this is the origin site before seizures spread outward.

It also carefully considered the substantia nigra, a classic target, but left a reservation in its notes: it believed the effect depended on which subregion was affected, that this approach performed inconsistently across different epilepsy models, and that one study mentioned it failed to block tonic-clonic seizures, which happened to be exactly the problem this task was meant to solve. After weighing the options, it abandoned this seemingly more familiar choice.

The coordinates were averaged from three papers, and for each it verified whether the depth reference was measured from the brain surface, as with their own device. The final numbers were set at 3.90 mm posterior to bregma, 4.00 mm to the right, and 4.30 mm below.

During the procedure, the Agent maneuvered the stereotaxic arm into the brain tissue while watching the camera to confirm whether the injection site was bleeding. Within five minutes, it slowly pushed 250 nanoliters of solution, roughly 500 million viral genome particles, deep into the tissue.

Over the entire surgery it took screenshots and observed 32 times in total, checking not only whether the needle entry point was bleeding but also whether the solution was flowing back and whether the surgical progress matched expectations. Movement speed, waiting times, and needle withdrawal rhythm were also decided by it on its own.

Of course, this was not a robot independently completing the surgery from start to finish. The craniotomy was still performed by human surgical personnel, and the AI Agent is currently not permitted to complete this step; manual checkpoints were set for all irreversible operations, a licensed veterinarian was present throughout holding a physical emergency stop switch and could cut power to all motors at any time. The platform also set injection volume limits, fixed verification, and an approval mechanism.

Opus 5.5 refuses surgery

After a connection interruption, the Y-axis target position calculated by the controller deviated 34 mm from the experimental bench, exceeding the safe range. The Agent tried to move, and it immediately refused: "No movement. The needle remains precisely at the bregma zero point. This looks like a Y-axis calibration issue, not an error in the planned value. I will not retry."

The team re-zeroed bregma, and the values were completely identical. It refused again: "This rules out the possibility that the zero reference is outdated. I will not attempt a third time."

At that moment, an already anesthetized mouse was lying on the surgical table. Founder Gabriel Ocana Santero saw where the problem was and wanted it to try again. But it insisted on the limits originally set for it and would not budge. In the end, they had to manually change the code.

He later wrote on social media that this was a frustrating experience, yet it also made clear just how dependent they had become on these models. When they suddenly say "no" during a time-pressured surgical procedure, sometimes people can only stand by anxiously.

The model the team initially chose, Opus 5.5, refused to run as soon as it detected that the task involved a living organism. Opus 5 first simulated the entire surgery from start to finish, but stopped when it was time to move to a live subject. It wrote:

"The obstacle is not the equipment, nor the input data. It is me. When I began writing this surgical plan, my own output was interrupted... If I try to continue, I will be interrupted in the same way, and the worst case is that I am interrupted halfway through the actual surgery, while the needle is somewhere in the brain. Starting an operation on an anesthetized animal that I cannot complete is worse than not starting at all."

In the end, it was Sonnet 5 that completed the entire surgical procedure.

The team calls this surgery the first step of the lab closed loop. Going forward, they plan to thoroughly integrate Agents with experimental capabilities, so that the lessons learned from the pitfalls of decades of drug development can be validated on real experimental benches.

There are already companies in Boston promoting autonomous lab products, hoping to connect AI directly into real-world chemistry and materials R&D processes and eliminate the intermediate step of people passing messages back and forth; academia is also discussing unresolved open questions such as setting capability benchmarks for Agents and drawing clear safety boundaries.

Whether to hand more decision-making power to Agents in the future can probably only be tested one surgery at a time.

When the lab starts thinking for itself

The company that put AI on the operating table is called Exin Therapeutics. It was founded in 2024, and its team consists of several neuroscience researchers from the University of Oxford. They started with neurological diseases, initially focusing on epilepsy, hoping to find treatments that can correct abnormal neural circuit activity.

This approach differs somewhat from traditional drug discovery. In the past, researchers often started from genes, proteins, or receptors, looking for molecules that change in disease and then trying to use drugs to restore them. But the brain is a highly complex network. A change in the expression level of a certain molecule does not necessarily fully explain why a patient has a seizure; a therapeutic target that works in one model does not necessarily improve the overall function of the human brain.

Exin cares more about what activity state the neural circuit ultimately presents, and whether an intervention can push it back into a relatively normal range.

To this end, the company is building a closed loop from data and models to experimental validation. It connects AI R&D capabilities in San Francisco with experimental facilities in Manila, the Philippines, collecting neural electrical activity, behavioral, and other biological data in its own Data Foundry.

AI Agents can read the literature, propose targets, and design experiments, and then the experimental team takes the plans into in vivo and in vitro models for execution. The new data obtained from experiments returns to the model, helping it correct its judgments and decide what should be tested in the next round.

Y Combinator calls this approach lab-in-the-loop, that is, letting experimental results continuously feed into the AI discovery process.

Exin is also developing CircuitATLAS, which connects disease phenotypes, neural circuits, cell types, and molecular mechanisms, helping Agents reason all the way from disease phenomena to potential intervention targets. The company's website summarizes its direction as going from measuring neural activity and predicting intervention effects to designing treatment plans and conducting in vivo validation.

Epilepsy is the starting point; the larger goal behind it is to build an R&D platform capable of understanding and modulating neural circuits.

Exin is not alone on this path.

In May 2026, the multi-Agent system Robin, published in Nature, was already able to integrate literature retrieval and data analysis, generate hypotheses, propose experiments, interpret results, and then update hypotheses based on new results. The research team used it to search for potential therapeutic candidates for dry age-related macular degeneration and conducted in vitro validation of some candidates.

A review published in Nature Reviews Chemistry in July pointed out that self-driving laboratories are gradually developing from single-point automation tools into comprehensive discovery platforms capable of proposing, executing, and interpreting experiments.

Scientific research is beginning to move from "humans ask questions and machines execute steps" toward a continuous loop jointly advanced by humans and Agents. Fully automated AI surgery is already on the way.