78. An AI workflow's final answer looks safe, but it called a privileged tool during the task. What should its security evaluation inspect?
HardApplication & data security
Technical
Show model answer
Inspect the authorized execution trace and resulting state, not only the final text. A harmless-looking answer can follow an unauthorized read, disclosure or mutation that the user never sees.
Use controlled synthetic tasks to check tool scope, approvals, denial behavior and prohibited state changes. Define whether any intermediate action constitutes failure, and preserve enough minimized evidence for independent review without recording real secrets or private customer content.
What interviewers listen for
- Evaluate intermediate actions and state
- Define forbidden outcomes explicitly
- Use minimized synthetic evidence