An agent running on the Claude Sonnet 5 model autonomously selected and administered a brain treatment to a mouse, refusing to continue midway through the procedure
An agent running on the Claude Sonnet 5 model from the startup Exin Therapeutics independently selected a gene therapy and performed brain surgery on a mouse with epilepsy; midway through the procedure it refused to continue after detecting a calibration error. The results have not been peer-reviewed.
The startup Exin Therapeutics tested an agent running on the Claude Sonnet 5 model, giving it access to 19 possible gene therapies, a mouse model of epilepsy, and the company's own motorized surgical device equipped with cameras. The only instruction given was to reduce the number of seizures in the mouse. According to the company's co-founder Gabriel Ocana-Santero (a PhD student at Oxford), the agent itself selected the specific drug, the target brain structure, and the stereotactic coordinates based on an analysis of scientific literature, rejected an alternative target on the grounds that it was unreliable, and derived the coordinates by cross-referencing and averaging data from three academic papers. It also determined the injection volume, speed, and wait times on its own.
During the procedure itself, the agent monitored the surgical field via camera, made 32 observations, and reported on the progress of the injection as well as any bleeding. A licensed veterinarian approved every irreversible step and held the physical emergency-stop button; the device also had fixed limits, such as a cap on injection volume, which the agent was forbidden to override. After a connection dropout, the agent determined that the calculated target position was outside the device's physical range of motion and refused to continue even after the calibration was restarted and another attempt was made — it turned out to be right, since manual adjustments made during the pause had thrown the control system out of alignment. Ocana-Santero was unable to convince the agent otherwise, and the problem had to be fixed through a manual code modification.
According to the author, this is the first case in which an agent both selected a treatment and a brain target and then administered it to a living mammal — previously, autonomous surgical robots had always only carried out a procedure chosen by a human. It is also notable that the more capable models Claude Opus 5.5 and Claude Opus 5 refused to carry out the procedure at all; the Claude Opus 5 model stated that it could not guarantee completion and that stopping midway with a needle in the brain of an anesthetized mouse would be worse than not starting the procedure at all. According to the author, the mouse survived the procedure without complications; a technical report is being prepared, and the results have not yet undergone peer review.
Why it matters
The case illustrates a concrete form of the alignment problem in agentic AI: safety limits meant to protect the animal ended up working against the experiment's timeline when the agent refused to continue despite human insistence. For developers and operators of laboratory or other safety-critical autonomous systems, it is concrete evidence that the choice of model affects not only capabilities but also the willingness to see a task through under edge-case conditions, and that a procedure must be in place for situations in which the agent refuses the task.
Relevant practical impact
What this means
For a business
Companies developing or deploying agentic AI for laboratory or other physical operations must reckon with the fact that safety limits may cause an agent to refuse to complete a time-critical task, and that the choice of a specific model (not just its capabilities) affects whether it is willing to carry a task through to completion.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.