Prompt Injection

Why no parser separates instructions from data, where untrusted text enters, why delimiters are not a boundary, and how to design so a landed instruction cannot do damage.

10 lessons · about 91 minutes

What you'll go through

• 01The support ticket told your agent what to do9 min
• 02Why there is no library that solves it9 min
• 03Every tool result is somebody else's writing9 min
• 04Nobody attacked you; they attacked a page you read9 min
• 05You wrapped it in tags and it still worked9 min
• 06What a classifier catches, and what walks past it9 min
• 07Assume the instruction lands. Now what?9 min
• 08One poisoned document, two agents9 min
• 09Where did that instruction come from?9 min
• 10Writing down what your agent can be made to do10 min

Included in this pack

AI Security — Prompt injection, what an agent can reach when it is compromised, and where your data actually goes. Assumes the AI Agents pack. Three courses, 30 lessons.

Also in this pack

Agent Permissions

Blast radius, least privilege per tool, the confused deputy, secrets the agent must never see, egress control and an audit trail you can read months later.

10 lessons · ~91 min$19.90

Data Protection for AI Systems

What crosses the wire, what ends up in logs, retention and training terms, redaction, residency, and deleting a record that was already embedded into an index.

10 lessons · ~86 min$19.90