AI/ML Security & Governance

Securing the AI systems your company is shipping right now — and proving each control to an auditor and a regulator

Every phase before this protected systems built by people; this phase protects systems that make decisions on their own. AI is now in production at most companies — chatbots on customer data, copilots with repo access, RAG over internal documents, agents that call tools and APIs — and it fails in ways the earlier controls were never designed for: a prompt that overrides the system instructions, a poisoned document that exfiltrates data through the model, an agent tricked into calling a tool it should never reach. This phase covers the 2026 AI threat landscape, the three governance frameworks a GRC engineer is now asked about by name (NIST AI RMF, ISO 42001, and the EU AI Act), prompt injection and its defenses, AI access control and LLM authorization, and the security of agentic AI, MCP, and RAG. The throughline is unchanged: translate an "secure the AI" requirement into a specific technical control and the exportable evidence that proves it operates.

~4 hours

4 modules

200 points

Module 7.2

AI Security Landscape + Governance Frameworks

The on-ramp to AI security: what actually goes wrong when AI ships to production in 2026, and the governance frameworks that now define the expectations. This module walks the AI attack surface — data, model, agent, and tools — with real incidents, then puts the three frameworks a GRC engineer is asked about side by side: NIST AI RMF (Govern/Map/Measure/Manage), ISO 42001 (the AI management system, the "ISO 27001 for AI"), and the EU AI Act with its risk tiers and obligations. You then classify eight AI use cases under the EU AI Act tiers — the exact judgement call that decides which controls a system must carry.

75 min

0 / 4

Module 7.3

Prompt Injection + Defenses

Prompt injection is the #1 risk on the OWASP LLM Top 10 and the AI-specific attack a GRC engineer will be asked about first. This module covers how it works — direct injection (instruction override, jailbreaks, prompt leaking) and the more dangerous indirect kind, where the payload is planted in a document, web page, or email the model reads — and then the defenses. There is no single fix, so the defense is layered: input validation, system-prompt hardening, output filtering, tool sandboxing, and treating retrieved content as untrusted. You finish by designing the defense-in-depth for a real LLM customer-support bot, mapping each attack to the controls that stop it.

75 min

0 / 4

Module 7.4

AI Access Controls + LLM Authorization

The defenses in the last module kept circling back to one fix: scope what the AI can reach. This module makes that the whole subject — authorization applied to AI. It covers the four questions behind every AI call (which identity, which model, which data, which action), the identity types you must tell apart (human user, service account, and autonomous agent), and RBAC/ABAC/ReBAC applied to models and data under least privilege and deny-by-default. Then it covers OAuth for AI: delegated versus service identity, the on-behalf-of flow, scoped short-lived tokens, and tool/MCP authorization — the controls that stop an AI from becoming a confused deputy with standing access to everything. You finish by designing an access matrix for six personas.

70 min

0 / 4

Module 7.5

Agentic AI, RAG Security + Red Teaming

The last and hardest setting: AI that plans and chains tool calls on its own, retrieval that pulls data into the context at query time, and the red-teaming practice that tests it all. This module covers the agentic threat model — excessive agency, tool misuse, runaway loops, goal and memory manipulation — plus MCP security and tool sandboxing; RAG access controls (retrieval filtered by the caller's permissions, provenance, corpus poisoning); and AI red-teaming basics. It ties every earlier thread together: prompt injection becomes tool actions, access control becomes what keeps an agent contained, governance becomes what you must prove. You finish by designing the security controls for an agentic AI assistant — a posture and guardrails for each thing it can do.

75 min

0 / 4