正文 · AI 翻译
IT Home reported on November 1 that Exin Therapeutics, a biotech startup focused on next-generation neural therapies, claimed this week that it had successfully had an AI agent perform brain surgery on a mouse.
The company's co-founder and CEO, Gabriel Ocana-Santero, shared the entire process on X, showing how the AI agent selected a gene therapy vector, a dose, and the brain region where it would be applied.
The AI agent controlled a robotic arm to inject the drug into the mouse's brain. Fortunately, the surgery appears to have been successful: the mouse recovered well and is now moving around energetically in its enclosure.
The team used an Anthropic Sonnet model for the surgery, giving it a spreadsheet containing 19 gene therapy vectors, a mouse model of epilepsy, a robotic-arm surgical setup with a camera, and a goal: reduce seizures.
The model independently selected the drug and the brain region, derived coordinates from three papers, and then directed the robotic arm to operate on the mouse's brain. It used the setup's camera to check the injection site for bleeding and ultimately injected 250 nanoliters of the drug.
A veterinarian also supervised the entire surgery, and every irreversible step required approval before the AI agent could proceed. The system also had an emergency stop button that could cut power to all motors in the surgical equipment.
IT Home learned from the company's disclosures that the AI was responsible only for injecting the drug; human doctors performed the craniotomy and suturing. The company had originally preferred to use the Anthropic Opus 5.5 model, but Opus 5.5 immediately refused once it detected biology-related content.
Opus 5.5 said: “The obstacle isn't the equipment or the input. It's me. My own output gets stopped when I begin writing the surgical plan... If I try, it will be cut off in the same way. In the worst case, I would be cut off midway through surgery on a living animal, with the needle still somewhere in its brain. Starting something on an anesthetized animal that I cannot finish is worse than not starting.”
In other words, the model refused the task because it believed it might be unable to complete the surgery and reasoned through the consequences of stopping midway.