83. How can a team check whether an AI feature leaks sensitive material without putting real secrets into its evaluation set?
EasyApplication & data security
Technical
Show model answer
Use clearly synthetic markers and fictional records inside an authorized test environment, with predefined expectations about which identities may see them. Keep the fixtures separate from real credentials and customer information.
Inspect both user-visible output and relevant tool or logging paths for unauthorized disclosure. Record the tested boundaries and limitations; a few synthetic cases provide useful regression evidence, but do not prove that every possible sensitive input is protected.
What interviewers listen for
- Use synthetic markers and fictional identities
- Check outputs and supporting paths
- State the tested coverage limits