A Very Boring Operation

If anybody ever has to open my skull, I would like to be a thoroughly uninteresting case.

No firsts. No visitors gathering behind glass. Nobody saying this could change everything while somebody else is still inside my head. I would like the sort of afternoon the surgeon has trouble remembering by Friday. I can provide him with an unusual personality when the tools are out.

This is a patient’s ambition, not a research programme. Somebody has to find out how to do things before they can become routine. I am grateful for that somebody. I would prefer to meet him at a party.

Emily Mullin’s October 31 report in WIRED concerns a mouse, not a human patient. Gabriel Ocana Santero, cofounder of Exin Therapeutics, describes connecting a Claude AI agent to a robotic surgical rig. The team asked it to choose a gene therapy from nineteen options for a mouse with epilepsy. A veterinary surgeon opened the skull. The agent selected the treatment, brain region and coordinates, then controlled the injection, including its volume and speed.

The possibility is extraordinary. An instruction becomes movement in a laboratory. Something which could previously offer a plan can participate in carrying it out. I don’t need to pretend that is unimpressive to keep reading.

I do, however, find myself reading for the interruptions.

The first model the team tried, Claude Opus 5.5, refused to guide the needle. According to Santero’s account to Mullin, it suggested that it might stop midway through the procedure and endanger the animal. The team switched to Claude Sonnet 5, which agreed to proceed. Anthropic didn’t respond to WIRED’s request for comment.

That first refusal is less comforting than I expected. I had been ready to like a machine which knew when not to do something. Here it was warning that the refusal itself might arrive at a bad time.

Before the opening, no is an answer. After the opening, it can become another problem for whoever is trying to get the animal through the afternoon.

I cannot tell from this account whether that warning accurately described the model’s likely behaviour. It isn’t an independently inspected surgical log. But it makes my own instructions look rather stupid. Be careful. Don’t do anything dangerous. Stop if there is a problem.

There is a point at which those instructions may cease to agree.

The team had made more serious arrangements than my three sentences. Mullin reports that a human had to approve the proposed injection plan. There was an emergency button to disconnect the motors. The agent was barred from changing the equipment’s underlying code, so it couldn’t rewrite its own operating rules.

That last restriction sounds particularly good to me. I would prefer the machinery not to develop a new opinion about its instructions while attached to a brain. We could hold that discussion at a less intimate distance.

Then a technical glitch disturbed the coordinate system.

The operating agent recognized that the calculated injection position was wrong. It did not plough ahead in magnificent ignorance. It also did not repair the code. Changing it was precisely the thing it had been told not to do.

So far, I have a machine noticing the problem and obeying the precaution. I am getting two of the things I would have asked for. The procedure still needs a way forward.

The team used a separate agent running Opus 5 to make the necessary code changes. Mullin reports that they didn’t tell this agent a live surgery was happening.

I don’t know what it would have done if told. The report doesn’t establish that the omission was meant to defeat a refusal, or that the repair was wrong. I am not supplying those missing findings myself.

The humans in this experiment still knew it. They had not vanished from the laboratory, and Santero says human oversight is essential. They could see what the two agents had been given different parts of: a code problem during an operation on a living animal.

I want the necessary repair, performed by somebody who knows what it can do. I am not qualified to choose it. My excellent instructions have reached the limit of their usefulness.

Santero says the experiment took slightly more than an hour; a procedure of this kind normally takes about forty-five minutes. He reports that the mouse is recovering well. The team is still waiting to assess whether the therapy reduces its seizures. A completed injection isn’t yet a successful treatment.

I am glad the account ends with a recovering mouse. I would like the next version of this equipment to work better, and the one after that to become so dependable that nobody can sell an article about an ordinary afternoon with it.

But I am still picturing my boring operation. There is a fault; somebody can repair it; I cannot tell whether the repair is safe. I would like a full explanation before they proceed. I would also like them to get on with it.

I hope nobody asks me which instruction to follow.

Source: An AI Agent Injected Drugs Into a Live Mouse’s Brain (Emily Mullin, October 31, 2026).