What happens when the product meets people who want it to misbehave, and when the content it processes contains instructions. Most of this is product work rather than security work, because the decisions are about what the system is allowed to do rather than about how to defend a boundary.
| # | Lesson | The question it answers |
|---|---|---|
| 01 | Prompt Injection as a Product Problem | Why is this not only a security issue, and what does product own? |
| 02 | The Guardrail Stack | Input, output and action guardrails: what does each actually catch? |
| 03 | Jailbreaks, Abuse and Rate Limits | How will this be abused in week one, and what is the ladder? |
| 04 | Redaction and Data Boundaries | What must never reach the model, and how do you prove it did not? |
| 05 | Content Policy for Generative Output | How do you write a policy, enforce it, and handle appeals? |
| 06 | Over Refusal | When is a model saying no a defect rather than a safeguard? |
| 07 | Untrusted Content in the Context Window | How do you handle text that arrived from somebody else? |
| 08 | Action Permissions and Confirmation | What should never happen without a person saying yes? |
| 09 | Transparency Duties and the AI Act | What are you obliged to disclose, as of today, and to whom? |
| 10 | Incident Response for an AI Feature | What is the runbook when the model says something it should not have? |
| 11 | Red Teaming Your Own Product | How do you find the failure before a user posts it? |
| 12 | Worked: A Guardrail Design | One real feature, with the stack and the escalation path |