3/5/26 The Cloud Security Alliance (CSA) released a paper on using Zero Trust for LLM environments highlighting how perimeter-era trust breaks down when prompts, tokens, integrations, and agents can move data across layers. For CISOs, the CSA piece provides an understanding of how the increase in the number of identities, integrations, and data paths impacts your threat landscape and guidance on how to uplift your Zero Trust strategy with mitigation controls.
Zero trust principles of ‘assume breach, trust nothing implicitly, continuously verify’ are applied to not just users, but models, data, agents, APIs, and every transaction. Taking a defense in depth approach and reframing the “LLM environment” as a layered ecosystem, the paper provides guidance across identity for humans and non-humans, network micro segmentation, continuous visibility, governance guardrails and critical input and output validation. In the context of zero trust, the paper describes threats beyond ‘prompt injection’ including data poisoning, model supply chain attacks, token theft, data exfiltration, shadow AI, and amplified blast radius in agentic AI.