What is tool leakage and how can you prevent it?

• A. When the agent uses tools only within the defined set; prevent with looser access controls.
• B. When the agent uses tools outside the defined set; prevent with magic.
• C. When the agent fails to log tool usage; prevent with better logs.
• D. When the agent uses tools outside the defined set; prevent with strict access controls and validation.

Answer: (D)

Tool leakage happens when an AI agent uses tools that are not in the approved, defined set. This can bypass safeguards and expose sensitive capabilities or data. The best way to prevent it is to enforce strict access controls and validation: maintain a whitelist of allowed tools, ensure every tool call goes through a controlled interface that enforces policy, and validate each request (which tool, with which parameters, and the scope of use). Adding runtime checks and a fail-safe, force-closed response if something isn’t allowed strengthens protection. Sandboxing and applying least-privilege to tool capabilities further reduce risk. While logging and auditing are useful for spotting leakage, they don’t prevent it on their own. The other options don’t address the core prevention: loosening controls invites leakage, and “magic” isn’t a real mechanism.