Delegation Escalation: The Hidden Security Risk in Multi-Agent AI

Enterprise AI systems face a new vulnerability called Delegation Escalation, where static service accounts in multi-agent chains create Confused Deputy attacks. The solution is OAuth 2.1 Token Exchange with actor claims, ensuring sub-agents only get the intersection of user and agent permissions. Non-human identities now outnumber human users 17-to-1 in cloud environments, making runtime auditability critical.