This website uses cookies to ensure you get the best experience.
Learn more.
A customer-facing assistant is tested only with polite ordinary questions. What additional evaluation is needed before release?
Published
- Only increase the number of ordinary greetings in the test set.
- Assume a commercial model cannot be influenced by malicious input.
- Disable all logging to ensure attacks leave no sensitive records.
- Run authorized adversarial tests for direct and indirect prompt attacks, data leakage, and unsafe tool use.
Comments — 0
Sign in to comment