AI Security 101 for Architects

Nine Parts in four moves. Select any move, Part or idea to open it.

Select a node. Dashed orange outlines mark key architecture connections. On a phone, scroll sideways.

HAS_MOVE HAS_MOVE HAS_MOVE HAS_MOVE The model is not the system Model + runtime context Hallucination as a property Hosted vs self-hosted I AI fundamentals an architect actually needs Capability ladder RAG ≠ authorization Tools and agents MCP II From business use case to AI capability Data can become instruction Dynamic trust boundaries Permitted execution space What hasn't changed III What actually changed in the architecture MOVE 1 Understand the system 7 threat families 12 threat questions Consequence-based risk Ten security principles IV AI security: from attack surface to threat model MOVE 2 Know the threats Identity ≠ authority Delegated authority Guardrails vs controls Control decision chain V From identity to runtime enforcement One request ≠ one call Agent-loop containment Fail-safe modes Cost per outcome VI Performance, scalability, reliability and cost MOVE 3 Control it Four assurance questions Behavioural regression Evidence ≠ logs Named owners VII Evaluation, observability, evidence and governance Five architecture planes Anti-patterns AEPVL review Checklists and canvas VIII The reference model and the AEPVL review Enterprise RAG assistant Tool-using agent Multi-agent / MCP Three failure questions IX Capstone projects MOVE 4 Prove it HAS_PART HAS_PART HAS_PART HAS_PART Opening: the book in ten minutes Closing and frameworks NEXT NEXT ASKS ? Central question AI Security 101 GUIDE