What problem does it solve? AI systems place language models in trust-bearing positions where they read attacker-controlled text, choose tool calls, and gate access to data, and conventional security review checklists do not cover the resulting risks like prompt injection, exfiltration through rendering layers, and agent permission bypass. This Skill provides structured, evidence-driven review checklists for every surface where a model sits in the trust path. ## Core Features & Use Cases - Surface-based review routing: Identifies which of six surfaces a target has — LLM application, agent runtime, MCP server, AI/ML supply chain, skill-shipping repository, or model-assisted vulnerability analysis — and applies the matching checklist from dedicated reference files. - Structural control verification: Enforces the rule that a control existing only in a prompt is not a control, checking harness-enforced permissions, sandboxing, egress proxies, credential isolation, and rail-enforced spend limits instead. - Restricted vulnerability analysis profile: Ships a fail-closed JSON profile template and a standard-library Python validator that must pass before any model-assisted source analysis begins. - Use Case: When reviewing an agent that reads GitHub issues and runs shell commands, use this Skill to map its indirect-injection paths, verify its sandbox and egress controls, and produce findings with severity, confidence, evidence, and remediation per the shared reporting contract. ## Quick Start Ask the agent to perform a security review of your LLM application or MCP server using the security-ai skill and report findings with severity and evidence.
