How to test a chatbot before launch
Pressure-test customer-facing chat before users teach it new tricks.
Run adversarial scans on the inputs real users will try, not just your demo script.
What breaks in chatbots
Users embed instructions in questions. Support bots leak internal guidance. Off-topic requests drift into unsafe advice.
Happy-path demos rarely include hostile phrasing or role-play tricks.
Start with these attack packs
Prompt injection, sensitive context exposure, and hallucination pressure cover most launch risks.
Add boundary bypass if you have strict scope limits.
Use the Release Gate
Let findings and policy drive a clear Ready, Monitor, Fix, or Blocked call.
Retest after fixes. A green demo is not proof.