Startup says Claude chose and performed a gene-therapy brain injection in a mouse

What the company says happened

Exin Therapeutics, a US biotech startup, says it let Anthropic's Claude Sonnet 5 take a central role in a gene-therapy brain injection in a mouse with epilepsy. According to the company, the AI was involved in the treatment plan, the choice of brain region, the injection coordinates and the control of the robotic arms. As technews.tw reported, the team calls the setup an "in vivo lab in the loop", meaning an AI wired directly into lab hardware that works on a living animal.

The claim comes from the company itself, mainly through a thread by CEO and co-founder Gabriel Ocana-Santero, which he wrote on the spot and described as brief. He promised more considered thoughts later. We have not seen independent confirmation.

How the setup worked

The researchers gave the AI four things: a spreadsheet listing 19 gene therapy vectors, a mouse model of epilepsy, a motorised surgery platform with cameras, and a single goal of reducing seizures. A vector is the carrier that delivers a gene-based treatment into cells.

From there, the company says, the AI decided which therapy to use and which brain region to target. It worked out the injection coordinates from three research papers. It then steered a stereotaxic rig, a robotic frame that places a needle at precise positions in the brain, and injected 250 nanolitres of solution. A nanolitre is a billionth of a litre. While it worked, it watched the entry point through the rig's camera for signs of bleeding.

Why this differs from a surgical robot

Ocana-Santero drew a line between this and earlier automated surgery systems. In those, he said, a human typically decides what the operation involves and the machine carries it out. Here the AI decided what to give and where to put it, then did the injection.

That shift, from a tool that executes instructions to an agent that chooses the action, is the real novelty. It is also why the claim draws attention and scepticism at the same time.

The humans and the safeguards

The team stresses that the AI did not run everything. A veterinarian performed the skull-opening step, and every irreversible action needed human approval. The vet also held a physical kill switch. The team also barred the AI from changing some low-level control code.

The run was not smooth. When the equipment's coordinate system had a calibration problem, the AI refused to keep going on a setup it judged to be wrong. The researchers ended up fixing the code by hand. In this telling, the refusal was a safety feature, though it also shows the system depended on people stepping in.

Other Claude models said no

The team also reported how other models behaved. Claude Opus 5.5 reportedly refused to carry out the task once it recognised it as a biological experiment. The earlier Claude Opus 5 could complete the planning on paper but would not do the live surgery. Its reasoning, as the team relayed it, was that being forced to stop partway, with instruments still inside an anaesthetised animal's brain, would be more dangerous than finishing.

This suggests that different versions of the same family of models handle risky physical tasks quite differently. It is also second-hand information, reported by the company.

What nobody knows yet

Ocana-Santero says the mouse recovered well after the operation. But the work has not been peer reviewed, and whether the treatment actually reduces seizures has yet to be observed. One mouse recovering says little about whether the AI's choices were good ones.

Outside observers have raised ethical and accountability concerns. Even with human oversight, they argue, handing high-stakes medical decisions to an AI calls for stricter safety rules. If something goes wrong, it is not obvious who is responsible: the lab, the vet, the software maker or the model.