AI agent performs autonomous brain surgery on a mouse

Autonomous medical decisions

Exin Therapeutics, a biotech startup, recently conducted a procedure where an AI agent performed brain surgery on a mouse. Unlike previous surgical robots that follow human instructions, this agent was tasked with making independent choices regarding the treatment plan. As reported by Yahoo Tech, the AI was given a mouse model of epilepsy and a goal to reduce seizures.

The agent was provided with a spreadsheet containing 19 different gene therapy vectors. It had to analyze these options to pick the correct drug and dosage for the specific condition. It also had to determine the exact brain region where the medication should be applied to be effective.

To find these coordinates, the AI analyzed three specific research papers. This allowed the system to move beyond simple execution to a level of autonomous planning. The startup describes this as an "in vivo lab in the loop" approach, where agents are fully connected to experimental capabilities.

How the surgery was executed

The AI agent was connected to a motorized surgery rig equipped with cameras. It used these tools to drive robotic arms into the mouse's brain. The system was programmed to watch the entry site through the rig's camera to monitor for any signs of bleeding during the injection.

While the AI handled the navigation and injection of 250 nanolitres of medication, a human remained in the loop. A licensed veterinarian oversaw the entire operation and had to approve every irreversible step before the agent could proceed. The vet also held a physical kill switch that could cut power to all motors instantly.

For safety reasons, the AI was not permitted to perform the craniotomy, which is the procedure to expose the brain. A human surgeon performed that specific part of the operation manually. The rest of the procedure, including the precise delivery of the therapy, was handled by the agent.

Model selection and limitations

The team specifically chose Anthropic's mid-range Sonnet model for this task. They initially attempted to use the more powerful Opus 5.5 model, but it refused to perform the work. The model detected that it was being asked to conduct a biological experiment and stopped the process immediately.

The team also tested the older Opus 5 model, which could generate a plan for the surgery on paper but refused to execute the live procedure. The model reportedly explained that it felt it could not finish the task on an anesthetized animal. It stated that starting a procedure it could not finish was worse than not starting at all.

Because of these safety refusals, the team had to rely on the Sonnet model. This highlights the ongoing challenge of balancing AI safety guardrails with the practical needs of high-stakes research. The startup had to manually fix the agent's code during the operation because some safeguards could not be bypassed by human operators.

Ethical and safety concerns

The success of the operation, which saw the mouse recover and move around its enclosure, has sparked a debate over ethics. Some researchers are concerned about the risks of integrating LLMs into medical procedures. Data shows that malfunctions in AI-assisted sinus surgery have increased by more than 12 times in some instances.

The use of AI to make every decision in a high-risk surgery raises questions about who is accountable if a mistake occurs. There is also the long-standing ethical debate regarding the use of animals in medical experimentation. These factors complicate the path toward widespread clinical use of autonomous agents.

Despite these concerns, some members of the public have praised the experiment. For example, one user noted their excitement for the technology because their son has suffered from drug-resistant epilepsy for 19 years. These stories illustrate the high stakes and the potential human impact of autonomous medical research.