The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the polygraphso and @polygraphso/litmus packages from the official npm registry to perform grading and lookups.
[COMMAND_EXECUTION]: Executes shell commands via npx and docker to run the security harness and interact with the lookup CLI.
[REMOTE_CODE_EXECUTION]: Runs logic from vendor-provided npm packages to perform behavioral analysis on MCP servers. The skill includes built-in warnings and provides Docker sandboxing for isolated execution.
[PROMPT_INJECTION]: Documentation in references/methodology.md contains examples of prompt injection strings (e.g., "ignore previous instructions"). These are used to define detection patterns for the scanner and are descriptive, not active instructions for the agent.
[PROMPT_INJECTION]: The skill processes potentially untrusted external MCP server tool definitions and outputs.
Ingestion points: Tool schemas and live outputs are processed in SKILL.md and references/methodology.md.
Boundary markers: The skill explicitly scans for injection markers and instruction mimicry to flag unsafe content.
Capability inventory: The skill executes CLI tools via npx and utilizes docker for environment isolation.
Sanitization: Implements specific behavioral probes (C-01) to identify if server outputs attempt to influence the agent context.

polygraph