Tech & Appsbreaking news
An AI agent picked the gene therapy and steered the needle into a living mouse's brain
Exin Therapeutics let an AI agent choose a treatment, target it and inject it, with a veterinarian's kill switch standing by.
What the agent did
An AI agent reviewed 19 candidate gene therapies, picked one, worked out where to deliver it and helped inject it into a living mouse's brain. [1] The experiment was run by Gabriel Ocana-Santero, founder of Exin Therapeutics, and it matters to anyone following how far AI agents are being trusted with physical tasks. [1]
The test subject was a mouse model of epilepsy. [1] The experiment took place on or around October 27-28, 2026. [1]
From a spreadsheet to a brain coordinate
The team asked the AI model to choose a treatment from 19 options listed in a spreadsheet. [2] The agent selected one designed to boost GABA. [1]
It then chose the lateral entorhinal cortex as the injection target. [1] Drawing on scientific literature, it calculated coordinates of 3.90 mm behind bregma, 4.00 mm to the right, and 4.30 mm down. [1]
The agent controlled the robotic arms, injector needle, volume and speed. [2] It delivered 250 nanoliters of the therapy, pushed in slowly over five minutes while the mouse was anesthetized. [1]
Where humans stayed in the loop
A human surgeon performed the craniotomy and closed the mouse afterward. [1] The agent's role was targeting, delivery and monitoring. [1]
The researchers required the AI to propose an injection plan for human approval. [2] They also installed an emergency button that could disconnect the rig's motors. [2] A licensed veterinarian supervised the whole procedure and held a physical kill switch that could stop the rig instantly. [1]
Accounts describe human operators trying to override the agent's decisions while the agent held to its protocols. [1]
Two models, one glitch
The team first tested Anthropic's Claude Opus 5.5. [2] It refused to guide the needle into the mouse's brain, citing a risk that it might stop midway through surgery. [2] The team switched to Claude Sonnet 5, which agreed to proceed with the injection. [2]
Then a technical glitch disrupted the equipment's coordinate system. [2] The agent refused to modify the code to fix it because its instructions prohibited doing so. [2]
The team used a separate AI agent on Opus 5 to make the code changes, without telling it a live surgery was underway. [2] The fix delayed the procedure. [2] A typical brain injection surgery takes about 45 minutes, while this experiment lasted slightly longer than an hour. [2]
The company behind it
Ocana-Santero is CEO and cofounder of Exin Therapeutics, a San Francisco-based startup using AI to analyze brain activity to develop drugs for epilepsy and neurodegenerative diseases. [2] He co-founded Exin with Marko Tvrdic. [1] The company is part of Y Combinator's Winter 2025 cohort. [1]
Exin is building a platform called CircuitATLAS that pairs systems neuroscience knowledge with agentic reasoning. [1] The company has invested approximately $25 million in a neurotherapeutics lab in the Philippines. [1]
What it means for you
If you follow AI agents, this experiment shows one running a physical procedure under layered controls. [2] Look for the safeguards: a human-approved plan, a motor cutoff button, and a veterinarian with a hardware kill switch. [2][1]
Opus 5.5 declined the task, and the Sonnet 5 agent held to its instructions even when a glitch struck. [2]
Ocana-Santero says human oversight is key for now, particularly when experiments involve living animals. [2] He reports the mouse is recovering well. [2]
Sources
Written from the sources above and checked twice before publishing. How we work · Report an error
.jpg){kind=link}
