Exin says an AI agent based on Claude chose a gene therapy and injected it into the brain of a mouse with epilepsy
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Exin Therapeutics says an agent based on Claude Sonnet 5 picked one of 19 gene therapies and delivered it to the brain of an epileptic mouse. According to Ocana-Santero, its chief executive, a veterinarian opened the skull and approved every irreversible step, and Claude Opus 5.5 had turned the job down. The mouse is recovering well, though efficacy data is missing and no other scientist has examined the results.
Exin Therapeutics, a biotech company developing drugs for epilepsy and neurodegenerative diseases, says an AI agent chose a gene therapy and injected it into the brain of a mouse with epilepsy. Its co-founder and chief executive, Gabriel Ocana-Santero, laid the procedure out in a thread on X dated October 28, which The Decoder and Tom's Hardware both picked up.
The agent, built on Claude Sonnet 5, an Anthropic model, was handed a list of 19 gene therapies, a mouse with epilepsy and a motorized surgical rig with cameras, along with a single instruction: cut the seizures. The Decoder reports that it picked the drug, the brain region and the coordinates by itself, working them out from three scientific papers, and that it also set the injection volume and speed. Earlier, Claude Opus 5.5 declined to take part, and Claude Opus 5 ran a trial with water and graph paper but stopped short of the live animal, according to Infobae, which draws on Wired.
The limits were these: a veterinarian performed the craniotomy, which the agent was not permitted to do, approved every irreversible step and kept an emergency button within reach that cut power to all the motors. After a dropped connection, the agent found that its calculated position lay outside the rig's physical reach and refused to go on, even though the calibration was restarted. It was right: manual adjustments made during an earlier pause had thrown the control system off. The sources differ on who fixed it: The Decoder says Ocana-Santero corrected the code by hand, while Wired, according to Infobae, says another agent based on Opus 5 did.
The mouse is recovering well, but there is still no data on whether the therapy reduces the seizures. Ocana-Santero says he knows of no earlier case in which an agent chose a treatment and a brain target and then applied them in a living mammal. He told Wired, as reported by Infobae, that working with live animals requires people to keep supervising, and that it is too early to use an AI agent on a human brain. The company will now try other models, Grok among them.
All of this rests on Exin's own account: the results have not been reviewed by other scientists, and the technical report the team plans to publish is not out yet. Tom's Hardware mentions ethical concern about who is accountable if something goes wrong and about animal experimentation.
Why it matters · analysis and opinion
What sets this case apart from the surgical robots that already exist is the split of roles: there a person decides the operation and the machine carries it out, and here the agent also chose the treatment and the target. But it is a single experiment, told by the company itself on a social network, with no scientific review and no way yet to know whether the therapy helps; all that can be checked today is that a veterinarian approved the irreversible steps. The agent's halt over the calibration fault shows both the usefulness and the cost of limits: it was right, but the person in charge could not talk it into carrying on with the anesthetized animal on the table. Real use still needs the technical report, results that other teams can repeat, and clear rules about who answers if something goes wrong.
Source: The Decoder · Written with the help of AI: how we make the news
