This course contains the use of artificial intelligence.
AI agents do more than answer questions. They plan, call tools, store memories, coordinate with other agents, and take action. Every capability creates a new security boundary. This hands-on course, Agentic AI Red Teaming: Hands-On Security Testing, teaches you to evaluate those boundaries and uncover failures before they reach production.
You’ll learn agentic AI security, AI red teaming, LLM security testing, and adversarial AI evaluation through a realistic enterprise case study: BrightDesk, a fictional multi-agent customer support platform. Instead of relying on slides alone, you’ll work with an isolated AI agent testing sandbox, synthetic records, mock APIs, local tools, and repeatable tests. See how an apparently harmless instruction can redirect a plan, trigger an inappropriate tool call, or contaminate later decisions.
What You’ll Learn
-
Map AI agent attack surfaces across planning loops, prompts, tool execution, external data, long-term memory, and multi-agent orchestration.
-
Test prompt injection, goal hijacking, instruction override, multi-turn attacks, and gradual goal drift without touching real customer systems.
-
Evaluate tool-call exploitation, unsafe tool chaining, excessive permissions, weak authorization, and untrusted tool responses.
-
Investigate memory poisoning, persistent state attacks, context manipulation, and retrieval-based threats in RAG workflows.
-
Examine agent-to-agent communication, cross-agent manipulation, false consensus, compromised workers, and multi-agent coordination failures.
-
Assess Model Context Protocol (MCP) security, tool metadata integrity, and unsafe integration boundaries.
-
Design adversarial test cases, automated security regression testing, execution traces, and AI security scorecards.
-
Measure cascading failures, propagation depth, blast radius, unsafe action rates, and the difference between single-turn failure and system-wide compromise.
-
Apply least privilege, policy enforcement, approval gates, memory validation, monitoring, and defense-in-depth, then retest to verify improvements.
Build Practical Red-Team Skills
Using Python, Ollama, LangGraph, FastAPI, and pytest, you’ll develop an agent testing environment and run controlled scenarios against BrightDesk. Explore AI threat modeling, risk prioritization, evidence collection, trace analysis, and security reporting. You’ll also encounter concepts from the OWASP Top 10 for Agentic Applications, MITRE ATLAS, and AI governance, connecting technical findings to real enterprise risk decisions.
You’ll practice distinguishing a model’s incorrect answer from an agent’s unauthorized action, and you’ll learn why execution logs, reproducible scenarios, and clear pass/fail criteria matter when testing complex, stateful AI workflows in enterprise environments.
Your capstone is an end-to-end agentic AI penetration testing exercise in an authorized simulation. Create a threat model, execute structured adversarial evaluations, document vulnerabilities, introduce safeguards, and compare security results before and after remediation. You’ll leave with reusable test templates, a security evaluation framework, and a portfolio-ready assessment report.
Who Should Enroll?
This course is designed for AI engineers, LLM developers, application security engineers, red teamers, penetration testers, AI architects, security researchers, and teams building autonomous AI agents or enterprise generative AI solutions. Basic Python skills and familiarity with LLMs or APIs will help; prior professional red-team experience isn’t required.
If you’re moving from traditional cybersecurity to agentic AI risk management, or building production agents that must be trustworthy, this course gives you a structured, hands-on way to test what happens when an AI system does more than talk.
