🔍⚛️ One AI balked at brain surgery - another dove right in US biotech startup Exin Therapeutics that analyzes brain activity to develop drugs for epilepsy and neurodegenerative diseases decided to test whether AI could move from designing laboratory experiments to actually performing them, WIRED reports. Connected to a custom-built surgical rig with cameras, robotic arms and an injector, Anthropic’s Claude Opus 5.5 agent was tasked with carrying out a brain injection after a veterinary surgeon opened a live mouse’s skull. It then: 🔴 selected a treatment from 19 gene therapy candidates 🔴 identified the target brain region and calculated the injection coordinates 🔴 controlled the robotic arm, needle, injection volume and delivery speed However, what followed baffled researchers. After helping plan the operation, the AI agent refused to guide the needle into the animal’s brain, apparently flagging the risk of causing harm if the procedure went wrong. The researchers turned to Claude Sonnet 5 – it agreed to proceed. There was also a technical snag: when a glitch disrupted the equipment's coordinate system, the agent recognized the problem but refused to modify the underlying code because its instructions prohibited it. A separate AI agent on Opus 5 stepped in to fix the issue. Human safeguards remained in place, including approval of the injection plan and an emergency shutdown button. However, AI models are trained primarily on published scientific literature - the sanitized record of what worked. That means they may miss the failed experiments that never saw daylight, or the split-second decisions made under pressure when something starts to go wrong. The mouse is reportedly recovering well. 👍 Substack | Chat | @geopolitics_prime