A biotech startup says one of its AI agents has done something no surgical robot has done before: chosen the therapy, the dose, and the target region itself, and then driven the robotic arms that injected it into a living mammal's brain. Exin Therapeutics, a startup focused on next-generation neurotherapeutics, says its Claude-controlled surgery rig operated on a mouse model of epilepsy — and that the mouse recovered well.
The company's co-founder and CEO, Gabriel Ocana-Santero, shared the procedure in a thread on X on October 28, writing that the experiment took place the night before. The claim has not been peer-reviewed, and the company says a full write-up is planned for its Substack. Still, the account offers one of the most concrete glimpses yet of what labs call "in vivo lab in the loop" — agents fully connected to experimental capabilities rather than chat windows. For more stories like this, follow our latest AI developments coverage.
What the Agent Actually Decided
According to Ocana-Santero's thread, the team gave the agent 19 gene therapy vectors in a spreadsheet, a mouse model of epilepsy, access to a motorized surgery rig with cameras, and a single goal: reduce seizures. The agent, running on Anthropic's Sonnet model, picked the drug. It picked the brain region. It derived the injection coordinates from three research papers it was given access to. Then it drove the stereotaxic arms into the brain and injected 250 nanoliters of the therapy while watching the entry site through the rig's camera for signs of bleeding.
"Autonomous surgical robots exist and they are better at the mechanics than our bench rig," Ocana-Santero wrote. "But in every one of them, a human decided what operation to perform, and the robot executed it. Nobody told this agent what to give or where to put it... I am not aware of an agent previously choosing a treatment and a brain target and then delivering it into a living mammal."
Human Guardrails at Every Irreversible Step
The setup was not unsupervised. A licensed veterinarian gated every irreversible step and held a physical kill switch that would cut power to every motor in the rig. The vet also performed the craniotomy — exposing the brain — because the AI agent was not permitted to do that part. Several safeguards were hard-coded so the agent was banned from overriding them, and at one point during the operation the safeguards locked out the human operators too, forcing a manual fix to the agent's code mid-procedure, according to the thread.
Opus 5.5 Said No
Perhaps the most striking detail is which model refused the job. The team initially tried Anthropic's frontier Opus 5.5, which immediately declined to participate once it detected it was being used in a biological experiment. The older Opus 5 model, according to Ocana-Santero, was willing to plan the entire procedure on paper but refused to perform live surgery.
"The block isn't the rig and isn't the inputs. It's me," the model reportedly told the team, per Tom's Hardware. "My own output was stopped when I started writing the plan for this surgery... if I try, it gets cut off the same way, and the worst version of that is me cut off partway through a live procedure with a needle somewhere in the brain. Starting something I can't finish on an anesthetized animal is worse than not starting."
The startup ultimately used Sonnet, Anthropic's mid-range model, which did not raise the same objection.
A Claimed First, With Caveats
If verified, the experiment marks a genuine shift in what AI agents do in scientific research — from analyzing data to making treatment decisions and physically delivering them. But the evidence so far is a social media thread, not a paper. The procedure's success is self-reported, the video evidence has not been independently reviewed, and the ethical questions are substantial: accountability if an operation goes wrong, the welfare of animal subjects in automated experiments, and the precedent of letting a commercial model choose a biological intervention.
Researchers have previously documented risks when language models enter the surgical workflow — one study reported that adding an LLM to sinus surgery workflows saw malfunction rates rise more than twelvefold, as cited by Tom's Hardware. Exin's own framing is cautious on this point: the mouse, the company notes, recovered well and was recorded moving around its enclosure.
Either way, the experiment captures where the field is heading. Labs are wiring agents into physical equipment — pipettes, reactors, surgery rigs — and letting them close the loop between hypothesis and experiment. The question Exin's thread leaves open is not whether agents will run laboratory experiments, but how much autonomy to grant them, and which models will agree to take the job.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
