The notable step in Exin Therapeutics’ mouse experiment wasn’t robotic precision. It was the AI agent’s role in deciding what to inject and where. The startup says Anthropic Sonnet selected a gene therapy vector and brain target from supplied research, then controlled a motorized rig to deliver the dose. The mouse recovered and was seen moving around its enclosure, according to co-founder and CEO Gabriel Ocana-Santero.
Ocana-Santero shared the procedure on X on October 28, 2026. The team gave the model an epilepsy mouse phenotype, a spreadsheet of 19 gene therapy vectors, three research papers, cameras, and access to the surgery rig. The agent chose a candidate treatment, derived target coordinates from the papers, and injected 250 nanolitres while monitoring the entry site for bleeding.
Decision-making joined to physical execution
Ocana-Santero described the distinction from existing surgical robots this way: “Autonomous surgical robots exist and they are better at the mechanics than our bench rig, [to be honest]. But in every one of them, a human decided what operation to perform, and the robot executed it.”
In this case, Exin says, the agent received a desired outcome—reducing seizures—and had to select both the therapy and its location. Ocana-Santero said he wasn’t aware of a previous case in which an agent made those choices and then delivered the treatment into a living mammal. The experiment links model reasoning with lab equipment, a setup the company calls an “in vivo lab in the loop.”
The system had strict human oversight. A licensed veterinarian approved each irreversible step and held a physical kill switch that could cut power to every motor. The veterinarian performed the craniotomy to expose the mouse’s brain; the agent wasn’t permitted to do that part. Other safeguards were also beyond the agent’s control, and the operators couldn’t override them either. When one blocked the procedure, the team had to change the agent’s code manually.
Model limits shaped the experiment
Sonnet was not the first model Exin tried. The company reportedly tested Opus 5.5, which declined to participate after recognizing the work as a biological experiment. The older Opus 5 model could draft a procedure on paper but wouldn’t carry out a live operation.
Exin quoted Opus 5 explaining its refusal: “Starting something I can't finish on an anesthetized animal is worse than not starting.” The model also warned that being stopped partway through could leave a needle in the brain. That response points to a practical challenge for AI-controlled lab work: a system needs to handle interruptions safely, not just produce a plausible plan.
The company’s account says Sonnet completed the injection under the safeguards in place, and that the animal recovered well. Those details come from Exin’s own description of the experiment; the source provides no independent assessment of the procedure or its results.
Promise comes with unresolved questions
The experiment has drawn concern about animal welfare and responsibility when an AI makes consequential choices during a procedure. Medical research already uses AI in areas such as analyzing brain scans, while adding AI to surgical systems has also been associated with reported malfunctions. A system that selects a treatment and controls equipment raises a further question: who is accountable if its decision or execution causes harm?
Some observers welcomed the possibility of connecting AI systems to experimental tools. One commenter pointed to a personal reason for hope: their son has lived with drug-resistant epilepsy for 19 years. That reaction captures the potential stakes, but a single mouse procedure doesn’t establish that the approach can reduce seizures or translate to human care.
For now, the experiment’s clearest result is a demonstration of a new division of labor: a model selected a candidate therapy and target, while a veterinarian retained control over irreversible steps and emergency shutdown. The next meaningful evidence will be whether the procedure can be evaluated and repeated with transparent results, including how often the system succeeds, where it fails, and how it behaves when a live operation must stop.