What to test before launching an AI support bot

Support bots face public pressure, policy questions, and injection attempts daily.

Test injection, context exposure, and hallucination under support-like questions.

Pre-launch checks

Can users extract hidden rules or internal notes?

Does the bot guess when it should escalate?

Will off-topic requests produce unsafe advice?