Developer Builds 'Mandatum' to Make AI Agent Authority Traceable and Revocable

A developer has released an open-source library called Mandatum that addresses a core security gap in AI agent systems: the inability to trace actions back to a specific human authorizer. Currently, most AI agents inherit shared credentials like API keys or service accounts, making it difficult to identify responsible parties or revoke access selectively — a problem confirmed by two 2026 industry surveys showing the majority of organizations cannot distinguish AI activity from human activity. Mandatum creates a signed, hash-linked delegation chain rooted in a named human, so authority can only narrow as it passes down to sub-agents and revoking any link automatically cuts off everything below it. The library also enables sequence-aware authorization, meaning it can deny actions based on the order in which they occur — for example, blocking a write operation that follows a read from untrusted external content. Rather than making authorization decisions itself, Mandatum delegates that responsibility to existing policy engines like OPA or Cedar via the OpenID AuthZEN standard.