Gabriel Ocana-Santero, co-founder of startup Exin Therapeutics and a PhD student at Oxford University, stated on X: They provided an AI agent 19 with a table of gene therapy vectors, an epileptic model mouse, and an electric stereotaxic surgery table equipped with a camera, with the sole goal of "reducing seizures." The agent independently selected the drug and brain region, derived and verified injection coordinates from three papers, then operated a robotic arm to inject 250 nanoliters while monitoring for bleeding via the camera. Every irreversible step was approved by a licensed veterinarian, who held a physical emergency stop switch.
Confidence: Medium (60) · Unverified · Event time 2026-10-28
Reasoning
- Researchers claim this is the first time an AI agent both selected the treatment plan and brain region and delivered the treatment to a live mammal; previous surgical robots only executed plans chosen by humans.
- According to The Decoder, after a connection interruption during the procedure, the agent discovered that the calculated target position exceeded the device's travel range and refused to continue; recalibration did not change its judgment. It was later proven correct (previous manual adjustments had messed up the control system), but researchers could not persuade it to continue and had to manually modify the code. Researchers say this exposes the problem of relying on models in time-critical operations when the model says "no" mid-process.
- According to The Decoder, only Claude Sonnet 5 was willing to complete the live surgery from start to finish; both Claude Opus 5.5 and Opus 5 refused. Opus 5's reason was that it could not guarantee completion, and stopping with the needle left in the anesthetized animal's brain mid-procedure is worse than not starting.
Usage
- None. Process screenshots and descriptions can be viewed in the researcher's X post; the technical report has not yet been published.
Risks
- Currently, there are only the researcher's own X post and a media report, with no paper or technical report and no peer review; there are no results yet on efficacy (whether seizures were reduced).
- Only the first post in the X thread was read; details on model selection, refusal specifics, and 32 observations come from The Decoder's summary of the entire thread.