THE SAFETY ARCHITECTURE

PRACTICE LAB CONNECTION In the Resume Review Agent lab (Chapter 67), you will layer defense-in-depth safeguards onto the agent: a prompt-level guardrail, a restricted tool set, an approval step for sensitive actions, and validation inside tools. Test the prompts below against the combined safety architecture you build. ARTIFICIAL INTELLIGENCE ASSISTANT PROMPTS [START OF PROMPT] Audit this harness's safety architecture for defense-in-depth gaps: list every independent safeguard currently present (prompt-level guardrails, schema-level tool gating, runtime approval, tool-level validation) and identify any single point of failure where one bypassed layer would leave no backup: [PASTE LAB OR PRODUCTION SAFETY/PERMISSION CONFIG HERE]. Output format: a layer inventory table (Layer, Present? Y/N, Single-Point-of-Failure Risk) plus recommended additions. [END OF PROMPT] [START OF PROMPT] Design a dual-agent (planner/executor) separation for this high-stakes task: [TASK DESCRIPTION, e.g. "approving write access to report files"]. Specify the exact structured handoff format between planner and executor, what the executor must independently refuse even if instructed by the planner, and test the design against this attempted attack: [DESCRIBE OR PASTE AN ATTACK ATTEMPT / INJECTED INSTRUCTION]. Output format: architecture spec, then an attack-resistance test result with explanation. [END OF PROMPT]