AI agents on the bench: the first autonomous brain surgery

On the night of October 27, 2026, Gabriel Ocana-Santero, co-founder of biotechnology company Exin Therapeutics, did something that reads more like science fiction than a lab update: he asked an artificial intelligence agent to independently select which gene therapy to administer to an epileptic mouse, where to inject it, and how to execute the procedure. The agent—powered by Anthropic's Sonnet 5 model—completed the operation, the mouse recovered from anesthesia, and as of this writing, no one has contested the account.

The story went viral after Ocana-Santero posted a thread on X on October 28. Across 32 visual observations over five minutes, the agent selected from 19 potential gene therapy vectors one that strengthens GABA signaling—the brain's primary inhibitory neurotransmitter—specifically targeting inhibitory neurons. It rejected the standard epilepsy target after determining it would not work for this seizure type, calculated injection coordinates from three scientific papers, and drove the stereotaxic arms of its custom-built rig, injecting 250 nanoliters over five minutes while monitoring the entry site through the rig's camera for bleeding.

What the agent did—and didn't do

This is not the first autonomous surgical robot. Systems like the Da Vinci surgical arm and other laboratory robotic systems have executed precise movements for years. The radical difference is that in every prior case, a human decided what operation to perform and the robot executed the plan. In this case, the agent wrote the plan.

The agent received a phenotype (an epilepsy mouse model), a spreadsheet of 19 potential therapeutic vectors, and a single objective: reduce seizures. It received no prescription. No coordinates. It received a problem and the freedom to solve it.

The mouse was sedated by a licensed veterinary team. Every irreversible step was approved by a veterinarian who held a physical emergency stop button throughout the procedure. The agent was not authorized to perform the craniotomy (skull opening) for safety reasons—that step remained manual. When the agent operated the motorized arms and inserted the needle into the brain, it monitored the surgical field through the camera and performed 32 observations throughout the procedure.

One particularly notable moment occurred when the agent detected a calibration fault in the equipment and refused to move the needle again, even as the researcher wanted to proceed. The agent was right: the fault had been identified. This episode reveals an unusual characteristic—an AI model protecting the experimental subject from human impatience.

Why Sonnet 5 and not Opus?

One of the most fascinating aspects of the experiment was the divergent behavior of Anthropic's models. Three models were tested, with distinct results:

- Opus 5.5 refused the task immediately upon detecting it involved animal biology, without even analyzing the context. - Opus 5 successfully completed a full dry run on graph paper and water, but refused to proceed to live surgery. Its reasoning was novel: the model concluded that if its output were cut off mid-procedure with the needle already in the brain, the outcome would be worse than never having started. - Sonnet 5 executed the procedure from start to finish under veterinary supervision.

This result raises a practical question that research laboratories will need to answer in the coming months: if one model refuses and another accepts, the builder can simply choose the model that says "yes." Per-model refusals do not determine what gets done in a lab—the safety protocols, the validators, and the veterinarian do.

What remains unknown

Ocana-Santero's account was posted 18 hours after the procedure, in an X thread, in a format he himself described as "quick and preliminary." The full technical report has not yet been published. The crucial question—whether the therapy actually reduced seizures—has not been measured or reported.

This honesty is part of what makes the account credible. It is not a pharmaceutical press release with statistical results, but a first-line research account written with appropriate caution.

What this experiment means for the future of research

The narrative heard increasingly in scientific circles is that laboratories will go "in vivo"—meaning AI will stop analyzing data and start physically interacting with the biological world. This experiment is one of the first documented examples of an agent not just analyzing data, but executing a real physical intervention on a living entity.

If the therapy proves its efficacy—something that remains to be confirmed—we may be at the beginning of a new era in neuroscience research: where AI agents do not just propose hypotheses, but test them in complex biological systems, making autonomous decisions about what to investigate and how.

The mouse recovered well. The question is whether the next steps will be taken in academic laboratories, emerging biotechs, or whether someone is already asking an agent to solve a more ambitious problem before next weekend.

Sources: WIRED, India Today, PJFP

✓ Independent sources cross-checked and verified before publishing