The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a vulnerability surface by instructing the agent to ingest, evaluate, and harvest patterns from untrusted external sources, such as GitHub repositories, package registries, and curated lists. This data could contain hidden malicious instructions designed to influence the agent's behavior.
Ingestion points: Untrusted content enters the context via GitHub Search, GitHub Topics, Web Search, and multiple package registries (npm, PyPI, Crates.io, RubyGems, etc.).
Boundary markers: The skill lacks explicit technical delimiters (like XML tags or clear 'ignore instructions' warnings) for the ingested data. However, it does provide a conceptual 'Entry Protocol' and evaluation checklist to help the agent filter low-quality results.
Capability inventory: The skill empowers the agent to recommend the installation of dependencies (npm install, etc.) and to 'harvest' code patterns, which involves parsing and potentially executing logic derived from external, unverified code.
Sanitization: The skill mitigates risks by requiring the agent to adapt and rewrite harvested patterns for the local context and to verify license compatibility, acting as a form of manual sanitization and review.
[SAFE]: The skill references several well-known services and trusted sources for research purposes, including GitHub, npm, PyPI, and various language-specific registries. These references are used legitimately for searching documentation and code and do not involve suspicious execution patterns.

github-search