Transitive Policies for Agents | GLTHR.com, by Guillaume Lethuillier

The author argues that AI agent behavioral constraints can only survive delegation when an independent runtime preserves their authorization context, information-flow restrictions, and action history. Relying solely on prompt delimiters or tool permissions fails because untrusted content can be reinterpreted as trusted instructions, and downstream agents can strip away vital context. To maintain security, the runtime must mediate every action and enforce transitive policies outside the models. Open article →