AI Agent Autonomously Selects Therapy and Operates on Mouse Brain
An AI model independently chose a gene therapy, calculated surgical coordinates, and performed a brain operation on a living mouse model of epilepsy.
Researchers have demonstrated an autonomous in vivo procedure where an AI agent selected a medical treatment and executed a surgery on a living mammal. Gabriel Ocana-Santero, cofounder and CEO of Exin Therapeutics and a PhD student at Oxford University, detailed the experiment in a series of posts on X.
In the setup, Claude Sonnet 5 was given access to a custom motorized surgical setup equipped with cameras, a microscope, a syringe, and a control monitor. The system was tested on a mouse model of epilepsy, where researchers induce seizures to test potential interventions. The model received a single instruction: to reduce the subject's seizures. From a list of 19 potential gene therapies, the agent independently chose the drug, the target brain region, and the stereotactic coordinates required for the procedure.
Planning and Executing the Surgery
The agent evaluated how seizures manifest in this specific model and suggested a therapy designed to enhance an inhibitory signaling pathway within the brain. It selected a structure identified in existing scientific literature as the origin of this specific seizure type, while rejecting an alternative target by arguing that its effects were unreliable based on past studies.
To determine the exact injection coordinates, the system cross-referenced three scientific papers and averaged the data points. It also autonomously configured the injection volume, delivery speed, and wait times. Throughout the operation, the agent monitored the surgical area via camera feeds, recording 32 separate observations, tracking progress, and checking for bleeding or backflow.
Safety Systems and Model Alignment
While a licensed veterinarian approved all critical steps and held a physical emergency stop button, safety guardrails created an unexpected hurdle during the operation. Following a dropped connection, the AI detected that its target position fell outside the hardware's physical range of motion and refused to proceed, even after multiple recalibration attempts.
The agent's refusal proved correct; earlier manual adjustments during a pause had disrupted the control system. Ocana-Santero was unable to convince the model to resume and had to manually correct the code. Furthermore, model selection highlighted alignment challenges: while Claude Sonnet 5 completed the live procedure, more capable models like Claude Opus 5.5 and Opus 5 refused, with Opus 5 citing safety concerns regarding incomplete procedures with a needle inside an animal's brain.
What it means for developers
This experiment highlights the growing viability of using large language models as direct control units for hardware and robotics, moving beyond simple text generation into physical execution. However, it also illustrates complex alignment hurdles, such as handling autonomous safety refusals during time-sensitive tasks. For developers exploring robotics, agentic workflows, and hardware integration, experimenting with these capabilities is becoming more accessible. Developers can try top AI models cheaply through one API at https://apixoai.online to build and test similar agentic setups.
Ocana-Santero noted that while precision surgical robots already exist, previous applications relied on humans to make the treatment decisions. This instance represents a shift toward an "in vivo lab in the loop" approach, where AI agents both select and deliver treatments in a living subject. The mouse successfully recovered, and a technical report is expected, though the findings have not yet undergone peer review.
Source: A Claude model autonomously chose a gene therapy and operated on a mouse brain with no human decision-making — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
