In May 2026, NSA, DHS CISA, and their Five Eyes partners published “Careful Adoption of Agentic AI Services,” the most authoritative national-security-level guidance on agentic AI governance issued to date. The paper correctly identifies that agentic AI introduces structural and behavioral risks beyond traditional software, that privilege misconfiguration is catastrophic at agentic scale, and that accountability is among the hardest unsolved problems in AI deployment. These are accurate observations. They are not governance primitives. They are symptoms of a deeper structural failure that the guidance cannot reach because the conceptual vocabulary required to address it—formation layer, identity drift, authority-at-formation, pre-instantiation governance—does not exist within the institutional frameworks the guidance is bound to extend. The guidance is not wrong. It is incomplete in a specific and consequential way: it identifies agentic risk at the execution layer and proposes execution-layer mitigations, while the governance failure occurs at the formative layer—before the system runs anything.
Narnaiezzsshaa Truong (Mon,) studied this question.