Agentic AI risk isn't a model problem. It's an architecture problem.

This article outlines how agentic AI and LLM-based systems shift security risk from software components to the data and context they consume, increasing risks like prompt injection and poisoned retrieval-augmented generation. It compiles expert guidance to mitigate these risks — including least-privilege access, strict policy checks, data provenance, sandboxing, input sanitization, runtime monitoring, containment (kill switches, network isolation), and tamper-evident auditing — and argues that traditional AppSec controls alone are insufficient.