✳ Learning path / 3 lessons
AI application and agent security
Three intermediate lessons establish defense-in-depth for LLM applications and autonomous tool-calling agents: isolating untrusted user and RAG data from instructions (OWASP LLM01:2026), enforcing deterministic Policy Enforcement Points against Excessive Agency (LLM03:2026), and wiring out-of-band human oversight gates mapped to the NIST AI RMF.
Work through the lessons in order to build a complete engineering mental model, or jump straight to the topic you are designing or reviewing today.
OWASP GenAI LLM Top 10 2026 and NIST AI RMF 1.0 stay on their own explainer pages. Completion of G1 through G3 does not seal the LLM Top 10 or AI RMF.
01Prompt injection and untrusted retrieval (RAG): when data hijacks instructions
Large Language Models do not natively separate developer instructions from untrusted data in the context window. Learn how direct and indirect prompt injection exploit RAG pipelines and how to contain them.
Excessive agency: why a model sentence is never an authorization decision
Giving an AI agent broad tools or ambient credentials can let prompt injection or model errors cause unauthorized changes. Learn how to separate model proposals from independent policy and approval gates.
Human oversight gates and the NIST AI Risk Management Framework
A rubber-stamp "Approve" button does not make an AI system safe. Learn how to design meaningful Human-in-the-Loop (HITL) gates bound to immutable parameters and map controls to NIST AI RMF 1.0.
Test your engineering judgment
Every lesson closes with a 4-choice scenario check modeled on real architectural trade-offs. Ready for full hands-on simulations or certification drills?
Open the practice labs ↗ · Launch the CISSP practice simulator ↗