An agent is software that decides at run time which tools to call with which arguments, so the security question shifts from what the user can do to what the agent can be talked into doing. This course treats agents as identities with privileges. You start with agent anatomy, the OWASP Top 10 for Agentic Applications (ASI01 to ASI10, released December 2025) and the Five Eyes joint guidance Careful Adoption of Agentic AI Services (May 2026). You then design agent identity and authorization: workload identities, on-behalf-of access with scoped OAuth tokens, user-scoped authorization enforced outside the model, and least agency through allowlists, parameter constraints and budgets.
A full module covers the Model Context Protocol as of the 2026-07-28 specification: its trust model, OAuth-based authorization and audience-bound tokens, confused-deputy risks, tool poisoning and rug pulls, and how to vet, pin and host servers. You will design execution sandboxes (containers, gVisor, microVMs) with egress and secret controls, secure inter-agent communication including A2A, contain cascading failures and rogue agents, and lock down coding agents in developer environments. The course ends with audit-grade tracing, kill switches and a pre-production checklist. Prompt injection mechanics are covered in depth in sec-210. The Wirewalk Institute certificate attests to completion and assessment of this course; it is not endorsed by OWASP, MITRE, NIST or any vendor named here.