This paper argues that contemporary “model security” narratives in AI are a Trickster archetype at work: they displace the real governance problem by centering attention on the model artifact rather than the substrate that actually rules. Using the Trickster at the village gate as a framing device, it shows how cinematic threats—poisoned weights, backdoored graphs, spectacular misdetections—become a festival that reassures auditors while leaving the true sources of power in agentic systems ungoverned. In Mythos‑class, multi‑model infrastructures, the model is a mask, not the god: the behavioral substrate composes models, tools, agents, and contexts into emergent configurations that artifact‑layer scanning cannot see. The paper contrasts “shiny” artifact‑layer dangers with deeper physics dangers arising from unconstrained authority lineage, unclassified drift, broken interpretive continuity, and the absence of substrate‑level invariants. It concludes that securing the mask does not secure the system, and that effective AI governance must operate at the substrate level—where identity, authority, and behavior are constrained before any model or mask is invoked.
Narnaiezzsshaa Truong (Mon,) studied this question.