What problem does it solve? AI and agent systems introduce trust boundaries that traditional security reviews miss: prompt injection, retrieval poisoning, over-broad tool permissions, and unvalidated model output reaching interpreters. This Skill provides a structured, defensive review workflow to identify and remediate those AI-specific threats before they become incidents. ## Core Features & Use Cases - AI Boundary Mapping: Records models, prompts, retrieval sources, memory, tools, MCP servers, and downstream consumers to establish exactly where untrusted content crosses trust boundaries. - Threat-to-Control Analysis: Checks for direct and indirect prompt injection, confused-deputy tool use, cross-tenant memory leakage, and unvalidated model output, then maps each threat to the smallest relevant control. - Safe Test Matrix: Verifies defenses using synthetic fixtures and inert canaries, confirming that untrusted content is never promoted to instructions and that rejected actions fail safely. - Use Case: Before shipping an agent that reads user documents and calls internal APIs, run this review to confirm tool calls are schema-validated and authorized independently of the model, and that retrieved content cannot hijack the agent's instructions. ## Quick Start Review this agent system's prompt, retrieval pipeline, and tool permissions for prompt injection and trust-boundary failures, then report findings with evidence and recommended controls.
