CourseAI Engineering · 0622 minPremium
Safety
Assume the model will be persuaded, by a user, a document or a tool result, and design so that being persuaded is not enough to cause harm: a threat model, five layers against indirect injection, an output guard that checks text before it streams, identity, rate limits and spend caps that hold under concurrency.
Premium
Part of the AI Engineering course, on the Premium plan
Readers with Lifetime Access read it in full. Already bought it? Sign in from the top bar.
What’s inside
- 1. Threat model
- 2. Layer 1: strip what readers cannot see
- 3. Layer 2: scan retrieved text, then apply a policy
- 4. Layer 3: spotlighting
- 5. Layer 4: least privilege on a tainted turn
- 6. Layer 5: the prompt
- 7. The output guard
- 8. Identity, rate limits and budgets
- 9. Evaluating safety
- 10. Known residual risks
- What breaks in production
- Exercises
- Check yourself