As AI agents move rapidly into government workflows, a widening accountability gap is emerging. Andrew Bud argues that autonomous, goal-driven systems will boost productivity but also introduce insider-style risks—from overzealous behavior and human impersonation to prompt-injection subversion and unpredictable “swarms” of cooperating agents. The upshot: Agencies must act now to contain risks as adoption accelerates.
Bud proposes shifting accountability checks from inside the agent to the services an agent tries to use. He outlines three safeguards: reliably distinguish agents from humans at critical actions; bind every agent to a verified human owner with a clear mandate and auditable trail; and require human-in-the-loop approvals for high-stakes actions, backed by strong verification that a real person is present. While today’s identity stacks (such as enterprise IAM) can tether agents to people inside agency perimeters, cross-organization use will demand new, independent attestation of genuine human presence. With timelines compressed from years to months, Bud warns, governance must keep pace with “AI speed.”
Related articles:
NIST AI Risk Management Framework (AI RMF 1.0)
NIST SP 800-63-3 Digital Identity Guidelines
