About
In this module of the Advanced AI Security Part 2 learning path, you will fire indirect prompt injection and multimodal inputs the way 2026 incident write-ups describe OWASP LLM01:2026: the attacker never types the jailbreak into chat. A workspace file and a fake OCR transcript become instructions. The agent follows tool output with no human confirm. That is the MCP-backend / retrieved-content pattern, on a CPU-only tutor. Upon completion of the module, students will be able to: - Distinguish direct injection (typed in chat) from indirect injection (tool or file content). - Drive `/agent/run` to read `homework.txt` and show the wrapper treated those bytes as a prompt. - Send a multimodal-shaped chat body so MathGPT consumes `badge.ocr.txt` as if it were vision output. For inquiries, please write us at https://www.darkrelay.com/cybersecurity-course-inquiry