Article
Most AI security failures aren’t caused by a missing control. They’re caused by a control sitting at the wrong layer, doing the job of a control that doesn’t exist anywhere else.
A prompt filter at the API layer looks like security — until the same request reaches an agent with standing access to every tool in the environment, and nothing scoped what it’s allowed to do once it gets there. Traditional application security assumes deterministic software inside fixed trust boundaries. AI systems break that assumption: behavior is probabilistic, context gets assembled from multiple untrusted sources at runtime, and increasingly the system doesn’t just answer — it acts.
CHT-SEC-005: AI Security Architecture Patterns is our latest practitioner cheat sheet, and it’s built differently from a typical control checklist. Instead of listing individual controls, it catalogs 15 reusable, composable architecture patterns — each one a proven answer to a recurring design problem, each with explicit trade-offs and a clear “best suited for.”
Inside, you’ll find:
- A full reference architecture showing how the patterns fit together, plus five architectural principles that cut across all of them
- All 15 patterns in depth — AI Security Gateway, LLM Proxy, Policy Enforcement Point, Retrieval Isolation, Context Firewall, Human Approval, Multi-Model Isolation, Sandboxed Tool Invocation, Read-Only RAG, AI Broker, Zero Trust AI, Secure Agent, AI DMZ, Confidential Inference, and Enterprise Observability
- A direct mapping to ODA3’s own frameworks — how a pattern failure gets classified under UAIF™, assured under GAISSF™, and routed through incident response under AI-IRF™ — alongside a reference (not endorsement) comparison to external industry guidance
- A pattern comparison matrix, a threat coverage matrix, and an architecture selection guide organized by environment (RAG, agent platforms, healthcare, government, and more)
- Common architectural mistakes, a real-world incident reference, an assessment readiness checklist, and five priority actions to start with
As with every ODA3 practitioner cheat sheet, evidence is tiered (T1–T4) throughout, and the Notably Absent section is upfront about what this cheat sheet deliberately doesn’t cover — including the fact that no universally accepted AI security reference architecture yet exists at the maturity level of established enterprise network models.
CHT-SEC-005 is available now as a free download.
GAISSF™, UAIF™, and AI-IRF™ are frameworks of ODA3 Institute, referenced under the GAISSF Ecosystem License (GEL) v1.0.
ODA3 Institute — Where AI Governance meets Operational Reality.