• Rule: whatever protections you implement, you must test that they actually work — and test the entire application together: LLM plus guardrails, classifiers, extra steps, and data access. 4:21
  • Red-teaming: act as a malicious user and try to break your own system to verify it responds as expected. 4:38
  • Concrete method: build a dataset of tricky adversarial prompts designed to manipulate the LLM into generating something it shouldn't, then check the app consistently denies them. 4:38
  • Safety is ongoing: new attack types keep appearing, so re-test the app continuously while it stays in production. 4:55