The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The documentation identifies potential attack surfaces when agents ingest untrusted data from external tool outputs or peer agents.
Ingestion points: Described in references/mcp.md (resource reading) and references/a2a.md (message turns).
Boundary markers: The skill explicitly mandates treating all external output as untrusted input and recommends human-in-the-loop consent.
Capability inventory: The guidelines cover tool execution and programmatic code execution architectures.
Sanitization: The documentation provides reference implementations for authentication middleware and advises on output filtering and truncation.
[EXTERNAL_DOWNLOADS]: Fetches official utilities from well-known GitHub repositories.
Evidence: references/registry.md provides a command to download the mcp-publisher binary from the official Model Context Protocol organization on GitHub.
[DYNAMIC_EXECUTION]: Outlines secure architectures for runtime code generation and execution.
Evidence: references/mcp-scale.md describes 'code mode' where models generate scripts to process data, providing detailed requirements for network isolation, resource limits, and per-call authorization within sandboxes.

agent-interop