An AI Agent Chose a Gene Therapy and Injected It Into a Live Mouse's Brain
Exin Therapeutics says Claude Sonnet 5 picked the treatment, the brain target and the coordinates, then drove the surgical rig. Claude Opus models refused, and seizure results are still pending.
Published
Last updated / Fact-checked on
An AI agent running on Anthropic's Claude Sonnet 5 chose a gene therapy, a brain target and injection coordinates, then carried out the injection on a living mouse through a motorized surgical rig, according to Exin Therapeutics. Its cofounder and CEO, Gabriel Ocana-Santero, says it's the first time he knows of an AI agent choosing a treatment and delivering it into a living mammal. Whether the therapy works is not yet known.
Photo: Konstantin Kolosov / Pixabay
How the experiment worked
Exin is a San Francisco-based epilepsy and neurodegeneration drug startup. The team gave the agent 19 candidate gene therapies in a spreadsheet, a mouse model of epilepsy and control of a custom motorized rig with cameras. Its goal was to reduce the mouse's seizures.
A veterinary surgeon first performed the craniotomy, the opening in the skull, while the agent observed. The agent then picked a therapy that boosts an inhibitory signaling pathway and a brain structure that research identifies as the origin of this seizure type, according to The Decoder.
It derived injection coordinates from three scientific papers, and set the volume, speed and wait times. During the procedure it watched through a camera and made 32 observations, including whether it saw bleeding, Ocana-Santero said.
| Item | Detail |
|---|---|
| Model that operated | Claude Sonnet 5 |
| Choices | 19 candidate gene therapies |
| Human role | Vet did the craniotomy, approved each irreversible step and held an emergency stop |
| Models that refused | Claude Opus 5.5 and Opus 5 |
| Seizure results | Pending |
Which models refused
The team tried Anthropic's more capable models first. WIRED reported that Claude Opus 5.5 refused to guide the needle, suggesting it might stop midway and endanger the animal. The Decoder reported that 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 agreed to proceed. Anthropic did not respond to WIRED's request for comment.
Safeguards, and a refusal mid-surgery
A licensed veterinarian approved every irreversible step and held a physical emergency stop that cut power to all motors. The agent was barred from changing the equipment's code.
Those rules caused a problem. After a technical glitch disrupted the rig's coordinate system, the agent correctly found that its target was outside the device's range and refused to continue. It also refused to change the code itself, because its instructions forbade it.
The accounts differ on how the team got past this. The Decoder says Ocana-Santero fixed the code by hand. WIRED says the team used a separate agent on Opus 5 to make the change, without telling it a live surgery was under way. The procedure ran just over an hour, against about 45 minutes typically.
"Frustrating experience but also revealing how much we can come to depend on these models and what happens when they just say 'no' in the middle of a time-sensitive procedure," Ocana-Santero wrote.
What happens next
The mouse recovered well, Ocana-Santero said, according to The Decoder. The team is now waiting for the therapy to take effect before measuring its impact on seizures, WIRED reported. A technical report is expected, and the results have not been peer-reviewed.
WIRED said the team planned to test Grok and other models next. In an October 30 post, Ocana-Santero said three models had each treated a separate mouse: Claude Sonnet 5 targeted the entorhinal cortex, OpenAI's GPT-6.1 Sol the hippocampus and xAI's Grok 4.7 the somatosensory cortex, according to TokenPost and EconoTimes. He said it was "too early to say" whether one did better. WIRED did not report this comparison.
Why it matters
Surgical robots exist, but humans have always chosen the procedure. Here the agent chose too. Ocana-Santero calls this "in vivo lab in the loop."
What to watch:
- Seizure results for each treated mouse.
- Exin's technical report and any peer review.
More coverage is on our Artificial Intelligence hub.
How we reported this
We read The Decoder's and WIRED's reports in full. Most details come from Ocana-Santero, through those outlets and his posts on X; we have not seen data from the experiment. Where the outlets differ, we report both. The three-model comparison comes from his October 30 post as reported by TokenPost and EconoTimes and is not confirmed by WIRED.
Published and fact-checked October 31, 2026.
Sources
- The Decoder, "A Claude model autonomously chose a gene therapy and operated on a mouse brain with no human decision-making," Oct. 30, 2026: https://the-decoder.com/a-claude-model-autonomously-chose-a-gene-therapy-and-operated-on-a-mouse-brain-with-no-human-decision-making/
- WIRED, "An AI Agent Injected Drugs Into a Live Mouse's Brain," Oct. 31, 2026: https://www.wired.com/story/an-ai-agent-did-brain-surgery-on-live-mouse/
- TokenPost, "Exin CEO Says Three AI Models Directed Gene Therapy Injections in Mice": https://www.tokenpost.com/news/technology/45424
- EconoTimes, "Exin Compares Three AI-Guided Gene-Therapy Procedures in Mice": https://econotimes.com/Exin-Compares-Three-AI-Guided-Gene-Therapy-Procedures-in-Mice-1767626
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.