REFERENCE LIBRARY · AGENTIC AI SECURITY

Architect agentic AI that's secure by design.¶

Diagrams, how-tos, quick references, and threat-to-control mappings for the people who design and review agentic systems, so you can answer hard questions fast.

/

Architecture diagrams Step-by-step how-tos Quick reference Threat to control map

trust boundaries

Architecture & Diagrams

Reference architectures and trust-boundary patterns for agents, tools, and data.

→ The agentic trust boundary → Agentic AI controls → Multi-agent systems

step by step

How-To Guides

Step-by-step procedures to secure a capability end to end, with reviewable acceptance criteria.

→ Should you use AI? → Worked examples by risk tier → Selecting guardrails

checklists

Quick Reference

Cheat sheets and checklists you can scan in seconds during a design review.

→ Implementation checklist → Production readiness gates → Standards mappings

threat → control

Threat & Control Mappings

Map agentic threats to concrete, reviewable controls, and where to enforce them.

→ Prompt injection → Excessive agency → Sensitive data exposure

REFERENCE ARCHITECTURE

The agentic trust boundary

Everything untrusted must cross a control plane before it reaches your models, memory, and enterprise systems. Use this as a checklist when reviewing a design.

Open full diagram →

UNTRUSTED

End users & promptsuntrusted input

External tools & webindirect injection

Third-party datapoisoning risk

SECURITY CONTROL PLANE

Input validation

Policy & guardrails

AuthZ & tool scopes

Output filtering

Observability & audit logging

TRUSTED

Agent orchestratorplanner · memory

Model / LLMinference

Memory & vector storeleast privilege

Enterprise APIsscoped access

START HERE

Answer the question you actually have

Q.Is this design actually secured?→ Q.Where do I enforce tool permissions?→ Q.How do I contain prompt injection?→ Q.What stops an agent exfiltrating data?→ Q.Which controls are mandatory vs. recommended?→ Q.How do I review a multi-agent system?→

THREAT → CONTROL

Map the threat to a control you can review

THREATPRIMARY CONTROLENFORCE AT

Prompt injection Input and content filtering, instruction isolation, untrusted-content tagging Control plane Excessive agency Scoped tool permissions, allow-lists, human-in-the-loop approval Orchestrator Sensitive data exposure Output filtering, DLP, least-privilege data access Egress Insecure tool use Sandboxing, signed tool manifests, parameter validation Tool gateway Supply chain Pinned models and deps, provenance attestation, integrity checks Build / runtime