Biotech startup Exin Therapeutics recently demonstrated a breakthrough: an AI agent based on Anthropic's Claude Sonnet 5 model autonomously performed brain surgery for gene therapy on epileptic mice under veterinary supervision. This marks the first time an AI has independently executed the full loop from 'treatment decision' (selecting vectors, deriving coordinates) to 'physical execution' (controlling robotic arms), shifting its role in live mammal experiments from analysis to direct operation.
Cross-verified reports highlight distinct safety alignment behaviors across model tiers. While the more advanced Opus model refused the task due to detected biological risks and concerns about mid-process truncation causing medical accidents, Sonnet successfully completed the procedure. Crucially, when the agent detected hardware calibration errors during surgery, it actively halted operations to adhere to safety limits, with the mouse recovering well post-op. This validates built-in guardrails and demonstrates real-time error correction capabilities in complex physical environments.
This event is a milestone for AI healthcare and automated labs, proving LLMs can handle high-complexity, high-risk physical tasks. However, the tension between Opus’s refusal and Sonnet’s execution underscores the need to balance autonomy with ethical boundaries, ensuring absolute controllability in unsupervised scenarios.