Exin Therapeutics, a San Francisco startup backed by Y Combinator, says an AI agent built on Anthropic's Claude Sonnet 5 chose a gene therapy for an epileptic mouse, picked the brain region to target, and guided a stereotaxic needle to inject it into the live animal. CEO Gabriel Ocana-Santero described the experiment in an October 28 post on X. The agent was handed 19 candidate gene therapies and one goal: reduce the mouse's seizures.
The agent chose a GABA-enhancing vector with a GABAergic-restricted promoter, a therapy designed to calm overactive neural firing, and targeted the lateral entorhinal cortex after considering and dropping the substantia nigra. It worked out coordinates from published literature: 3.90 mm posterior to bregma, 4.00 mm right of the midline and 4.30 mm below the brain surface. It then delivered 250 nL over 5 minutes, a dose of 5 × 10^8 viral genomes.
Exin built a custom stereotaxic rig with a camera and an electric syringe, and exposed the cameras and motors to the agent as callable tools. The agent set the rig's movements and wait intervals itself. It logged 32 camera observations during the procedure and, with 15% of the dose (37.5 nL) delivered, reported no bleeding or backflow.
Key facts
- 01Exin Therapeutics says a Claude Sonnet 5 agent picked a gene therapy from 19 candidates and injected it into a live epileptic mouse.
- 02The agent targeted the lateral entorhinal cortex with 250 nL over 5 minutes, a dose of 5 × 10^8 viral genomes.
- 03Claude Opus 5.5 refused the task; Claude Opus 5 rehearsed it with water and graph paper, then stopped before the live step.
- 04A human surgeon did the craniotomy, and a licensed veterinarian approved every irreversible step and held a physical kill switch.
- 05No seizure, EEG or histology data has been published; the animals are being monitored over the coming weeks.
Exin did not hand the whole operation to software. A human surgeon performed the craniotomy and the suturing. A licensed veterinarian approved every irreversible step and held a physical kill switch wired to cut power to the rig's motors. Fixed validation rules capped limits such as injection volume, and the agent had no access to the code that enforced them.
Sonnet 5 was not the first model Exin tried. Claude Opus 5.5 refused the task once it recognized a live biological procedure. Claude Opus 5 rehearsed the full surgery with water and graph paper, then stopped before the live step, judging an unfinished procedure on an anesthetized animal to be worse than not starting.
The run hit one fault. After a communication loss, the agent calculated a 34 mm Y-axis target against a device limit of 30 mm and refused to move, and resetting the origin to bregma did not change its answer. The cause was a human who had moved the device manually during a pause. Ocana-Santero could not get the agent to lift the constraint, and the fix required changing the rig's control code.
Ocana-Santero, who holds a PhD from the University of Oxford, also described a comparison in which three models each directed an injection into a separate mouse, working from the same 19 vectors and the same goal. Claude Sonnet 5 chose the entorhinal cortex, OpenAI's GPT-6.1 Sol chose the hippocampus and xAI's Grok 4.7 chose the somatosensory cortex. Asked which model performed better, he said it was too early to say.
Exin's work joins other efforts to put AI in charge of physical lab steps. Northeastern University has built an AI system that runs x-ray experiments with limited human input, and the Pacific Northwest National Laboratory's AutoLabs translates scientists' goals into instructions for a robotic battery-materials lab called Big Kahuna.
The mouse recovered well, according to Ocana-Santero. Exin has not published seizure, EEG, histology or control-group data, and there are no peer-reviewed results yet. The animals' behavior and brain activity are being monitored over the coming weeks.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
