AI red teaming borrows the discipline of traditional red teaming but adds targets that are probabilistic, conversational and connected to tools. This course teaches you to run engagements that produce findings developers can fix and leaders can trust. It starts with scope: separating security testing from safety and misuse evaluation, obtaining written authorization, and setting rules of engagement for systems that may touch production data or take real actions. Planning is threat-informed, using MITRE ATLAS (v2026.09 at review, 16 tactics) and the OWASP Top 10 lists for LLM and agentic applications.
You will work through jailbreak families, indirect injection and exfiltration chains, retrieval manipulation and corpus poisoning, and tool and agent abuse. You will automate with PyRIT, garak and promptfoo, use attacker models and multi-turn strategies such as PAIR, TAP and Crescendo, and score results with classifiers, LLM judges and human review. The course extends beyond the prompt to model-level testing, APIs, serving infrastructure and multimodal inputs, then covers attack success rates with confidence intervals, actionable reporting and coordinated disclosure. It closes with continuous red teaming in CI, building the team, and a capstone engagement against a hypothetical agentic application. All techniques are taught for authorized testing of systems you are permitted to assess. The Wirewalk Institute certificate attests to completion and assessment of this course; it is not endorsed by OWASP, MITRE, NIST or any vendor named here.