Unexpressible, Not Filtered: A Structural Approach to Agent Safety
中文摘要
作者建议通过结构化设计使AI智能体无法表达不安全行为，而非单纯依靠过滤，以实现更本质的安全保障。
English Summary
Instead of filtering unsafe AI actions, the author proposes a structural approach that makes certain unsafe behaviors impossible to express, ensuring fundamental agent safety.
原文节选
Why I stopped trying to catch unsafe AI-agent actions and made a class of them impossible to express instead — and the honest limits of… Continue reading on Medium »