Module 1: The AI attack surface

• Understanding the architecture of LLMs and AI agents: We will explore how Large Language Models process tokens and how "Agents" interact with external tools and APIs.
• Identifying trust boundaries in AI-integrated applications: Adding AI expands an organization's attack surface. Developers must identify new trust boundaries, such as the prompt interfaces and data streams, just as they would scrutinize exposed endpoints in traditional applications.
• Similarities and differences in vulnerabilities: There are strong similarities between traditional vulnerabilities and AI security concerns, as both often stem from the core failure to "never trust user input". However, the primary difference is the blurred lines between functionality and data in the case of AI. Unlike traditional injection flaws (like SQLi) that have a clear syntactic boundary between code and data, AI models process instructions and user data together as natural language, meaning there is no strict syntactic separation.

Module 2: Basics of prompt injection & jailbreaking

• Mechanisms of direct prompt injection: We will look at how attackers craft malicious inputs that trick the model into ignoring its developer-provided instructions, which is the AI equivalent of failing to sanitize user input.
• Basic techniques for bypassing safety filters (Jailbreaking): Just as relying solely on traditional input filtering is insufficient, we will cover the foundational methods attackers use to sneak malicious prompts past AI safety guardrails.