Certified Agentic AI Security Expert (CAASE)TM

Certified Agentic AI Security Expert (CAASE) is designed for security professionals who already work in offensive or defensive security roles and want to specialize in AI agent security, specifically penetration testers, red teamers, cloud security professionals, application security engineers, IT security managers, and DevOps engineers. The course is rated Intermediate to Advanced, so it assumes an existing security background rather than teaching security fundamentals from scratch.

Students should be comfortable with basic Linux commands (such as ls, cd, and mkdir), and familiarity with a scripting language like Python, Golang, or Ruby is helpful but not required. Practical DevSecOps also recommends completing its Certified AI Security Professional (CAISP) course beforehand, since CAASE builds on foundational AI security concepts.

The course content totals 40 hours: 8 hours of video lectures and 32 hours of hands-on labs, and it can be taken as a 3-day instructor-led training or as 36 hours of self-paced training. Enrolled students get 3 years of access to the on-demand course material, along with 60 days of access to the live online lab environment.

The CAASE course is organized into seven chapters:

• The agentic AI architecture, threat landscape, and toolchain;
• Attacking and defending agent reasoning, goals, and prompt boundaries;
• Attacking and defending agent memory, context, and tool use;
• Attacking and defending computer-use, browser, and coding agents;
• Assessing and securing multi-agent systems and agent identity;
• Operationalizing agentic AI security across DevSecOps, governance, and the supply chain; and
• Threat modeling agentic AI systems and agent communication protocols (A2A, ACP, and UCP).

CAASE includes more than 30 guided exercises, such as building a minimal agent harness, prompt-injecting an agent’s planning loop through a tool response, poisoning an agent’s long-term memory or RAG source, chaining tool calls to escalate privilege, escaping a coding agent’s workspace sandbox, spoofing agent-to-agent messages to hijack a delegated task, generating an AI/Agent Bill of Materials (AIBOM), and signing a harness with Cosign. All labs run in a browser-based lab environment accessible from a laptop, desktop, or even a tablet.

 The CAASE exam is a 6-hour, task-based practical exam (not a multiple-choice test) in which candidates solve five hands-on challenges that mirror the attack-and-defense scenarios covered in the course, such as exploiting agentic AI weaknesses and using automated tools against them. After finishing the exam, candidates have 24 hours to submit a written exam report for evaluation.

Candidates need to score at least 80 out of 100 points (80%) across the exam’s five challenges to earn the CAASE certification. Points are awarded based on how complete each challenge solution is, and each student receives one exam attempt as part of the course enrollment.

CAASE focuses specifically on the agentic layer: the harness, planning loop, memory store, and tool-calling logic that sit on top of a model, rather than attacking the underlying LLM itself. It goes beyond single-turn prompt injection to cover multi-agent delegation abuse, confused-deputy attacks between agents, computer-use and coding-agent workspace escapes, agent-to-agent protocol attacks (A2A, ACP, UCP); and supply-chain risks from open-weight models and marketplace skills, topics that generic LLM security courses typically don’t address.