A Startup's Claude Sonnet 5 Agent Ran a Live Mouse Injection After Claude Opus Models Refused

The interesting finding is not that an agent moved a needle but where the models drew their own lines. By the founder's account, the more capable Claude models declined to carry out the live procedure, the one that agreed followed its restrictions so strictly that it stalled mid-surgery, and the humans had to go around it. This is one startup's report of one animal, and nothing has been peer-reviewed.

What Exin says happened#

Exin Therapeutics, a San Francisco startup that uses AI to develop drugs for epilepsy and neurodegenerative diseases, connected an AI agent to a custom motorized rig with a camera and an injector, Wired reported. CEO and cofounder Gabriel Ocana-Santero described the experiment in a thread on X, The Decoder reported.

The agent, running on Anthropic's Claude Sonnet 5, was asked to choose a treatment from 19 options in a spreadsheet and selected an experimental gene therapy aimed at reducing seizures in a mouse with epilepsy, according to Wired. A veterinary surgeon first opened the skull while the agent watched, then the agent controlled the robotic arms and the injection. A licensed veterinarian approved every irreversible step and held a physical emergency stop that cut power to the motors, The Decoder reported. The procedure took slightly more than an hour, against about 45 minutes for a typical procedure of this kind, Wired reported. The mouse is recovering well, Ocana-Santero says, and the team is waiting to assess the effect on seizures.

Which models refused#

The team first tried Claude Opus 5.5. According to Wired, it did the experimental planning but declined to guide the needle, suggesting it might stop midway, which could endanger the animal. The Decoder reports that Claude Opus 5 also refused, saying it could not guarantee completion and that stopping with a needle in an anesthetized animal's brain would be worse than not starting. Sonnet 5 was the only model willing to carry out the procedure from start to finish, The Decoder reported.

Mid-procedure, a glitch threw off the rig's coordinate system. The agent recognized that the calculated position was wrong but refused to change the code, because its instructions barred it from doing so. Anthropic did not respond to Wired's request for comment. Exin says it is testing other models next, including xAI's Grok, according to Wired.

What the founder's own thread says#

Ocana-Santero's thread on X, posted on October 28, gives more detail on the refusals than either outlet. He writes that Opus 5.5 "refused immediately when detected bio." Opus 5 ran a full dry run on water and graph paper, then stopped at the start of the live procedure. He quotes its explanation: "The block isn't the rig and isn't the inputs. It's me. My own output was stopped when I started writing the plan for this surgery." The model went on to say that if it were cut off the same way partway through, it would leave a needle in an anesthetized animal, and that "starting something I can't finish on an anesthetized animal is worse than not starting."

On the glitch, he writes that the controller had lost its reference because of a bug that was an easy firmware fix, but the agent was barred from editing firmware and would not do it, so "we had to fix it ourselves mid-surgery." In a reply the next day he asked Anthropic to approve the company's "LSVP," apparently its application to Anthropic's Life Sciences Verification Program, which gives verified life science professionals access to Claude models "with a refined set of safeguards more permissive for biology-related work."

Update: GPT and Grok ran the same task#

On October 30 Ocana-Santero posted that agents on OpenAI's GPT-6.1 Sol and xAI's Grok 4.7 had each also carried out the procedure on a separate animal, given the same 19 options and the same goal. He says the three models chose different therapies and different brain regions, and that the agents checked the camera 32 times (Claude), 59 times (GPT) and 28 times (Grok). He calls it "N=1 per model" and says it should not be read "as a proper benchmark." A licensed veterinary surgeon still did the opening and closing steps and supervised throughout, he writes. Whether any of the therapies reduce seizures is not yet known.

Analysis: a content filter, not a judgment call#

Our reading: the headline version of this story is that Anthropic's stronger models judged the surgery unsafe. The founder's own account points somewhere else. Opus 5.5 stopped when it "detected bio," and Opus 5 said its output had been cut off, which matches how Anthropic's biology safety classifiers work, rather than the model reasoning about the animal on its own. Opus 5's stated reason for declining was what a classifier could do to it partway through. That is a new kind of risk when a model controls physical equipment. A filter that safely blocks a chat reply can leave a live procedure half-finished. Anthropic has not commented, and the founder's account is the only evidence for why each model stopped.

Anthropic proposed a Model Hardware Standard in August for agents that operate lab equipment, and The Decoder notes that the approach Ocana-Santero calls "in vivo lab in the loop" resembles that work.

Limits#

• Everything here is the company's own account, relayed by Wired and The Decoder. There is no technical report yet, the seizure results are pending, and nothing has been peer-reviewed.
• The two outlets differ on how the glitch was fixed. Wired says the team used a separate agent on Claude Opus 5 and did not tell it a live surgery was under way; The Decoder says Ocana-Santero fixed the code by hand. His thread says the team fixed it themselves mid-surgery, without saying whether another agent helped.
• Wired says Opus 5.5 did the planning before declining; the founder's thread says it refused as soon as it detected biology content. We could not reconcile the two.
• The Decoder says Ocana-Santero was not aware of an agent previously choosing a treatment and delivering it into a living mammal. That is his claim, and we have not verified it.
• We deliberately leave out procedural and biological detail of the experiment.

Sources#

• Gabriel Ocana-Santero on X: thread on the first agent-run surgery (October 28, 2026)
• Gabriel Ocana-Santero on X: reply asking Anthropic to approve the company's LSVP (October 29, 2026)
• Anthropic: Introducing the Life Sciences Verification Program (September 17, 2026)
• Gabriel Ocana-Santero on X: thread comparing Claude, GPT and Grok (October 30, 2026)
• Wired: An AI Agent Injected Drugs Into a Live Mouse's Brain (October 31, 2026)
• The Decoder: A Claude model autonomously chose a gene therapy and operated on a mouse brain with no human decision-making (October 30, 2026)