THE ESSENTIALS IN 30 SECONDS

  • According to Ocana-Santero, an agent on Claude Sonnet 5 chose among 19 candidate gene therapies, selected a brain target, computed coordinates from published papers and controlled a motorized injection rig.
  • Humans performed the opening and closing of the skull, a veterinarian approved every irreversible step and held an emergency stop, and hard limits capped what the agent could do.
  • The work is a single-animal demonstration reported on X; no peer-reviewed results or efficacy data exist yet, and a technical report is expected.

Gabriel Ocana-Santero, cofounder and CEO of Exin Therapeutics and a PhD student at the University of Oxford, has described on X an experiment in which an AI agent running on Anthropic's Claude Sonnet 5 planned and carried out a gene therapy injection into the brain of a mouse with epilepsy. The account has drawn wide attention, but it rests so far on his own report.

What the agent did

According to The Decoder, the agent was given a single goal, reducing seizures, along with 19 candidate gene therapies, an epilepsy mouse model and a custom motorized surgical rig fitted with cameras. It picked a therapy meant to strengthen inhibitory signaling, chose the lateral entorhinal cortex as the target after rejecting a second option it judged unreliable, derived coordinates from three papers and set injection volume, speed and waiting times. It then moved the device and syringe to deliver the dose.

Humans remained involved throughout. XenoSpectrum and Startup Fortune report that human surgeons performed the craniotomy, and XenoSpectrum adds that they also closed the wound. A licensed veterinarian approved each irreversible action and held a physical emergency stop, and fixed limits on movement and injection volume could not be overridden by the agent.

Refusals and a calibration fault

At one point a connection dropped, and after manual adjustments made during a pause, the agent found its target lay outside the rig's range of motion. It refused to continue, even after recalibration, until Ocana-Santero fixed the control code by hand. The Decoder says the agent was right about the fault.

Other models reportedly declined the task. Accounts say Claude Opus 5.5 refused, and that Claude Opus 5 rehearsed the procedure but would not perform it on a live animal; the outlets differ on the stated reasons, and XenoSpectrum notes these were not controlled comparisons.

Open questions

The mouse reportedly recovered, but no data show whether seizures were reduced. XenoSpectrum points out that only one animal was used, with no control group and no confirmation that the therapy was expressed. The results are not peer-reviewed, and a technical report is expected. Ocana-Santero says he knows of no earlier case of an agent choosing a treatment and brain target and then delivering it to a living mammal, a claim that has not been independently confirmed.