Federal AI systems now must log who authorized what — and prove it survived the attack

A new framework lets government agencies track exactly which AI agent did what on whose orders, and verify that delegated actions stayed within bounds. This means when a federal AI system breaches policy or gets hacked, investigators can reconstruct the chain of command and pinpoint where authority leaked.

Right now, when one AI delegates work to another AI that uses a tool on behalf of a human user, nobody can answer who actually authorized it or where the permission broke down. This paper shows a way to make that chain auditable and tamper-evident. The practical effect: federal agencies gain genuine accountability for multi-agent AI decisions instead of just hoping the system behaves. When something goes wrong, they can see exactly whose judgment failed.