PROJECT THREE - TOOLS, THE MODEL CONTEXT PROTOCOL (MCP), AND SAFETY

CHAPTER SUMMARY Chapter 67 added outside tools and safeguards together. The Model Context Protocol lets a client use a server's tools through one shared standard, and the lab registers them in the same registry as native tools. Idempotency means repeating an action has the same effect as doing it once, and the repaired write_report now reports unchanged for identical text and refuses replacement without approved overwrite. Hooks give the harness checkpoints before and after tools, including a human approval step. Prompt injection hides orders in data, tool poisoning hides them in tool descriptions, and defense in depth stacks several imperfect layers so no single failure is a disaster. KEY TERMS REVIEW MCP: an open standard for connecting models to tools and data. MCP client and server: the program that uses tools and the program that offers them. Idempotency: the property that doing something twice has the same effect as doing it once. Hook: a checkpoint where the harness runs your code at a chosen moment. Approval hook: a hook that waits for a human yes or no. Prompt injection: hidden orders placed in data the agent reads. Tool poisoning: hidden orders placed inside a tool description. Untrusted content: anything from outside that must be treated as data and never as orders. Defense in depth: stacking several safeguards so no single one must be perfect. Security event: a log record about something relevant to safety.