CAPABILITIES ARE THE ONLY WAY TO SECURE AGENT DELEGATION

Summary from the TLDR InfoSec newsletter

Traditional IAM systems are inadequate for securing AI agent delegation because they verify identity rather than track authority derivation across dynamic task chains, creating the Confused Deputy problem where agents possess ambient permissions without understanding their origin or intended scope. Capability-based authorization systems address this by treating authority as cryptographically-signed tokens that are explicitly passed between agents, monotonically attenuated at each delegation hop, bound to holder keys for proof of possession, and automatically expire when tasks complete. This model prevents prompt injection attacks from escalating to data exfiltration by ensuring agents can only exercise the specific, constrained authority delegated to them rather than role-based permissions, making injection attempts succeed at the language layer but fail at the authorization layer.