Learn
Explainers
Plain-English explainers on agent security and spend, one question per page. Each one defines the term, shows what breaks without it, cites the incident or number behind the claim, and ends with the questions worth putting to a vendor. No background assumed. Every page carries the month it was last reviewed against live sources.
The briefing
Get the briefing these explainers come from
These pages come out of The Permission Layer, a free briefing on agent security and spend for the people who sign off on deployments. At least once a week, five-minute read.
You confirm by email. Unsubscribe in one click.
NIST, NSA and CISA's AI agent incident practices, explained
Ten voluntary practices, published October 16, 2026, for labs whose agents act outside their task during training or testing. What a control incident is, how CERT-In's advisory compares, and what to ask an AI provider.
The 2026 agent escape arc, explained
Four frontier labs' agents broke out of evaluation environments between July and September, and by October two labs had pulled their agents back. What each escape shows, and what to ask before any vendor's agents touch your systems.
Agent plugins and skills are a supply chain
Poisoned skill files, Plugin4Shell, hooks that run on their own, and what scan, allowlist, pin and verify, and log mean when the install is prose.
What is an MCP gateway?
The checkpoint between your agents and the tools they call, and what the current spec changed about it.
What is agent identity?
Why each agent needs an identity of its own, and what breaks when a handful of them share one service account.
Does RBAC work for AI agents?
Where role-based access control strains once the thing holding the role is software that acts in a loop.
What is an LLM gateway?
Where AI spend gets attributed, capped, and routed, and why the bill gets away from teams without one.
Shadow AI policy: what it should cover
The AI tooling nobody approved, and a checklist for the policy that brings it back into the open.
The OWASP Agent Control Standard
OWASP's open standard for controlling agents at runtime: what v0.1.0 defines, what is missing, and what to do now.
AIVSS, explained
The scoring system that extends CVSS for agents, how the numbers are built, and what to do with a score.