> cat sans/sec546.md

SANS SEC546Course Author & InstructorPillar: Securing AI

Securing Agentic AI

I am the author of SANS SEC546, a five-day, hands-on course for defenders who need to secure AI agents that plan, call tools, remember, and act on their own. The course is built around practical controls rather than theory: every section pairs the threat with labs that implement the defense.

5 Days

30 hours, instructor-led

19 Labs

Hands-on, defense-focused

30 CPEs

Continuing education credits

Advanced

In-person or virtual

Why I wrote SEC546

Agents break many of the assumptions traditional application security relies on. They take instructions from untrusted content, hold long-lived memory, chain tools together, delegate to other agents, and increasingly act on desktops, browsers, and physical devices. Security teams are being asked to defend these systems before most of the playbooks exist.

SEC546 distills what I have learned securing AI and software supply chains into a defense-first curriculum. It follows an agent from its first prompt to its last action: boundaries and identity, runtime operations, the MCP tool layer, multi-agent and computer-use ecosystems, and finally agents that touch the physical world. I also teach SANS SEC545: GenAI and LLM Application Security, and SEC546 is designed as the natural next step for teams moving from LLM applications to autonomous agents.

What you will be able to do

Model agentic AI threats and attack surfaces
Enforce input, output, identity, and permission boundaries
Defend against prompt injection, context poisoning, and tool tampering
Harden agent memory and runtime operations
Secure MCP flows, tool execution, and desktop agents
Protect multi-agent and browser-based agent ecosystems
Apply cyber-physical safeguards in a capstone exercise

> ls sections/

The five days at a glance

Summaries below are my overview of each section. See SANS.org for the official syllabus, schedule, and registration.

Day 1

Foundations of Agentic AI Security

Builds the mental model for why autonomous agents change the threat landscape. Students threat-model agentic systems, put boundaries around what flows into and out of an agent, defend against prompt injection, and apply secure development patterns with scoped identities and the principle of least agency.

Topics

  • ›Agentic AI risks and threat modeling
  • ›Input and output boundary enforcement
  • ›Prompt injection defense
  • ›Secure agent development patterns
  • ›Agent identity, permissions, and least agency

Labs

  • 1.1Hardening with NeMo Guardrails
  • 1.2Agent Goal Integrity Controls
  • 1.3Build Secure Agent Chain
  • 1.4Privilege Scoping & Identity Controls

Day 2

Agent Operations, Hardening, and MCP Defense

Moves from design to day-to-day operations. Covers protecting agent memory and context stores, spotting and safely shutting down rogue agents, building observability and governance into agent runtimes, and placing a policy-enforcing gateway in front of Model Context Protocol (MCP) traffic.

Topics

  • ›Securing agent memory and context stores
  • ›Detecting, containing, and isolating rogue agents
  • ›Observability, governance, and continuous defense
  • ›MCP gateway defense and policy

Labs

  • 2.1Memory Integrity Controls
  • 2.2Safe Agent Termination
  • 2.3Runtime Governance and Policy Enforcement
  • 2.4Deploying Defensive MCP Gateway

Day 3

Secure MCP, Desktop Agents and Runtime Defenses

Goes deeper into the tool layer. Students detect poisoned context and tampered tool responses, sandbox tool execution with egress controls, lock down coding and desktop agents running on developer machines, and track the provenance of agent dependencies, skills, and prompts with an AI bill of materials mindset.

Topics

  • ›MCP data integrity and context security
  • ›Agent tool execution sandboxing
  • ›Securing desktop agents
  • ›Agent supply chain and AIBOM defense

Labs

  • 3.1Detecting Context Poisoning and Tool Response Tampering
  • 3.2Tool Sandbox and Egress Controls
  • 3.3Securing OpenCode Agents
  • 3.4Dependency, Skill and Prompt Provenance

Day 4

Multi-Agent, Browser and Computer-Use Agent Security

Addresses systems where agents talk to other agents and act in browsers or on screens. Covers trust between agents over agent-to-agent (A2A) protocols, sandboxing computer-use actions, scoping delegated authorization through token exchange, and preventing data from leaking across agents and tasks.

Topics

  • ›Multi-agent A2A protocol defense
  • ›Securing browser and computer-use agents
  • ›Delegated agent authorization
  • ›Cross-agent data leakage defense

Labs

  • 4.1A2A Trust Chain Controls
  • 4.2Cua Action Sandboxing
  • 4.3Token Exchange and Scoping Controls
  • 4.4Task Contamination Isolation Controls

Day 5

Cyber-Physical Agent Security & Emerging Frontiers

Extends agent security into the physical world, where mistakes have real-world consequences. Covers safety controls and kill switches for robotic and IoT agents, confidential execution and attestation for agents, and emerging research areas, then ends with a live capture-the-flag defense capstone.

Topics

  • ›Physical-world agent safety
  • ›Emerging agentic security topics
  • ›Agent defense capstone

Labs

  • 5.1Robotic and IoT Kill-Switch Controls
  • 5.2Confidential Agent Execution and Attestation
  • 5.3Live Defense Operations CTF

Who should attend

  • ›AI/ML security engineers building or defending agentic systems
  • ›AI security specialists and AI incident responders
  • ›Security operations and blue team leads
  • ›Cloud security engineers running agent deployments

Recommended background

  • ›Comfortable reading, modifying, and running Python
  • ›GenAI basics: LLMs, prompting, RAG, and tool calling
  • ›Application security fundamentals (injection, authN, authZ)
  • ›Command-line and cloud/API familiarity (API keys, env vars, JSON)

Take SEC546

SEC546 is offered by the SANS Institute. Course dates, formats, and registration are handled on SANS.org. Everything else on this site stays free.