Safety in Tool Calling

1. What is the primary principle behind treating tool requests from models as untrusted?

They are generated by a trusted process They could be manipulated or incorrect They rely on user authentication They follow a standardized format

2. Which of the following is a recommended practice for executing tool requests?

Execute without validation to save time Allow all tools by default Validate arguments against the schema and business rules Only check user permissions before execution

3. In the context of tool calling, what should happen when a model proposes a destructive action?

It should be executed automatically It requires human approval before execution It should be logged for future reference It should be ignored if deemed unsafe

4. What does a tiered policy in tool calling imply?

Tools should only have one level of access Actions are categorized by their level of risk All tools should be accessible at all times Tiered policies are not necessary for safety

Continue: When Not to Use Tools → Review this lesson Retake quiz