27. An AI assistant can read documents and send external messages. What security boundaries would you require?
HardRisk & governance
Technical
Show model answer
Separate untrusted content from action authority and restrict the tools, data, recipients, and operations available to the assistant. A document that tells the assistant to send confidential material is not authorization from the user.
Require appropriate confirmation for consequential actions, validate destinations and outputs, and retain privacy-safe audit evidence. Evaluate misuse and failure scenarios before expansion; fluent model output does not establish correctness or permission.
What interviewers listen for
- Content is not authority
- Least privilege for tools and data
- Validate consequential actions