An AI agent steered a needle into a live mouse’s brain after Claude Opus refused

An AI agent has chosen a treatment for a living animal, worked out where in its brain to put it, and then driven the robotic arms that delivered it. Exin Therapeutics, a San Francisco startup that develops drugs for epilepsy and neurodegenerative disease, connected a Claude agent to a motorized surgery rig with cameras, handed it an epileptic mouse and gave it one goal: reduce the seizures.

The detail people keep repeating is who said no. Anthropic‘s top model, Claude Opus, refused to put the needle in. The team switched to the cheaper Claude Sonnet, which went ahead. For anyone watching AI agents move off the screen and into labs, warehouses and cars, the experiment shows a model making the medical choices as well as the movements, and shows that models from the same company draw the line in different places.

According to a thread posted by Exin’s chief executive and co-founder, Gabriel Ocana-Santero, the agent received a spreadsheet of 19 gene therapy vectors, three research papers and control of the rig. It picked the therapy and the brain region, derived the coordinates from the papers, then drove the stereotaxic arms, the motorized frame that holds a needle steady in three dimensions, and injected 250 nanolitres while watching the entry point through the rig’s camera for 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.” His point is the reverse here: “Nobody told this agent what to give or where to put it.”

The refusal came first. The team started with Claude Opus 5.5, Anthropic’s most capable model, and it would not guide the needle into the brain, warning it might stop partway through the operation and leave the animal in danger, as Wired reported. “Starting something I can’t finish on an anesthetized animal is worse than not starting,” one of the Opus models told the team, according to Tom’s Hardware. Claude Sonnet 5 agreed to proceed. Anthropic did not respond to Wired’s request for comment.

Humans stayed in the loop at every step that could not be undone. A veterinary surgeon performed the craniotomy, the opening of the skull, while the agent watched; the agent is not yet allowed to do that part. The vet approved each irreversible step, the agent had to submit its injection plan for sign-off, and a physical kill switch could cut power to every motor on the rig. The agent was also locked out of the equipment’s code so it could not rewrite its own rules.

Those guardrails caused the main hiccup. A glitch scrambled the rig’s coordinate system, and the agent spotted that its calculated position was wrong but refused to edit the code, because its instructions forbade it. The team brought in a second agent, running the older Claude Opus 5, to make the fix, and did not tell it a live surgery was underway. A procedure that normally takes about 45 minutes ran a little past an hour.

The bigger caveats are about evidence. This is one mouse, described by the company itself on social media, with no peer-reviewed paper behind it. Nobody yet knows whether the therapy the agent picked works: the team is waiting for it to take effect before measuring the seizures. Tom’s Hardware notes the open question of who answers for a mistake when software makes every decision, on top of the old arguments about animal research. Nothing here is a product; there is no service for other labs to buy, in any country.

Ocana-Santero posted the thread on October 28 and promised a full write-up on Substack; Wired and Tom’s Hardware reported the experiment on October 31. Exin says it will test xAI’s Grok and other models next to compare them, and is working on agent-controlled craniotomies, though it wants to be sure they are safe before trying one on a live animal. Letting an agent near a human brain, Ocana-Santero told Wired, is far too soon.

The mouse, he says, is recovering well. Whether its seizures ease is now down to a therapy that a piece of software picked off a spreadsheet.

Tags: Anthropic, AI agents, Claude, robotics, Claude Opus 5.5, Exin Therapeutics