AI agent security is evolving at two speeds: most organizations still need foundational runtime enforcement (visibility, consistent controls, action-level enforcement before consequential actions), while Google's 'Beyond Zero' paper describes a future of dynamic, context-aware enterprise security decisions. The author argues that identity and least-privilege alone are insufficient — authentication proves who is acting, but not whether a specific action is appropriate. The confused-deputy problem is especially dangerous for agents that operate autonomously with delegated credentials. A practical starting framework is outlined: map one consequential workflow, place enforcement points before each risky action, standardize baseline controls (prompt injection, PII, rate limits), and only then layer in richer contextual reasoning. Arcjet is positioned as a tool providing this floor-level runtime enforcement today.
