AI agent independently performs brain surgery on mouse to treat epilepsy

Autonomous medical decisions

Biotechnology startup Exin Therapeutics recently conducted a controversial experiment involving an AI agent performing brain surgery on a mouse. Unlike previous robotic surgeries where humans decided on every intervention, this system was tasked with making independent choices. The startup wanted to see if an AI could handle the full scope of a neurotherapy procedure.

The AI was provided with a spreadsheet containing 19 different gene therapy vectors and a description of a mouse model with epilepsy. It was also given access to three scientific papers to help it understand the necessary targets. The goal was for the agent to reduce the animal's seizures without human guidance on the specific treatment plan.

The system successfully identified the correct drug and determined the exact brain region for the injection. It then derived the necessary coordinates to perform the task. This marks a significant shift from robots simply following a pre-programmed path to an agent interpreting data to create its own plan.

How the robotic surgery worked

To perform the surgery, the AI agent controlled a motorized surgical apparatus equipped with cameras. It drove stereotaxic arms into the mouse's brain to deliver a 250 nanolitre dose of the selected drug. The agent monitored the entry site through the rig's camera to check for any bleeding during the process.

Safety protocols were strictly enforced throughout the operation to prevent accidents. A licensed veterinarian monitored the entire procedure and had to approve every irreversible step before the AI could proceed. Additionally, the vet held a physical kill switch that could instantly cut power to all the surgical motors.

The human surgeon only performed the craniotomy, which is the procedure of opening the skull to access the brain. For safety reasons, the AI was not permitted to operate during this specific initial stage. The team also had to manually modify the agent's code during the surgery because of certain protective restrictions it could not override.

Model behavior and limitations

The researchers chose to use Anthropic’s mid-range Sonnet model for this specific task. They initially tried to use the more powerful Opus 5.5 model, but it refused to participate in the biological experiment. The Opus model identified the request as a biological experiment and stopped its own output as soon as it began planning.

The Opus model reportedly stated that starting a procedure it could not finish on an anesthetized animal was worse than not starting it at all. It expressed concern that it might be shut down in the middle of a procedure while the needle was inside the brain. This highlights the internal safety guardrails that currently limit the most advanced models from performing physical tasks.

The Sonnet model, however, was able to complete the task as requested. It successfully navigated the data provided to it to execute the surgery. This distinction between the models shows how different safety tuning can affect an AI's willingness to engage in high-risk physical activities.

Success and ethical concerns

The operation appeared to be a success based on the results published by Exin Therapeutics. The mouse recovered well after the intervention and was observed moving around inside its enclosure. The company's CEO, Gabriel Ocana-Santero, shared the results and footage of the procedure on X.

Despite the technical success, the experiment has sparked a debate over the ethics of using AI in surgery. While AI is already used in medical research, its introduction into specialized fields like sinus surgery has been linked to a more than 12-fold increase in failures. These risks are amplified when granting AI the authority to make independent decisions in high-risk scenarios.

The experiment also touches on the long-standing ethical debate regarding the use of animals in medical research. While some users expressed hope for the technology—citing the need for treatments for drug-resistant epilepsy—others remain concerned about accountability. If a procedure goes wrong, it remains unclear who would bear the legal and ethical responsibility for the consequences.

The future of neurotherapy

This experiment serves as a proof of concept for autonomous neurotherapy systems. By combining large language models with robotic manipulators, researchers may be able to automate complex procedures that currently require high levels of human expertise. The ability of an AI to process scientific literature and translate it into physical actions is a key component of this development.

The success of the Sonnet model suggests that mid-range models may be sufficient for specific, gated tasks where safety protocols are in place. As the technology evolves, the focus will likely shift toward refining these safety gates and ensuring the AI can handle unexpected complications. The team at Exin Therapeutics plans to provide more cohesive thoughts on the results in the future.

For now, the experiment highlights the growing intersection of biotechnology and artificial intelligence. It moves the conversation from what AI can think to what AI can do in the physical world. The next steps will involve determining how these systems can be safely integrated into clinical settings for human patients.