The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted web content from search results and user-provided URLs, creating a surface for indirect prompt injection.\n
Ingestion points: Untrusted content enters via the search and read commands in scripts/gemini_search.js.\n
Boundary markers: The instructions in SKILL.md and system prompts in the helper script explicitly warn the agent to treat this data as untrusted and not as instructions.\n
Capability inventory: Capabilities are limited to network requests to the Gemini API and external URLs for content retrieval.\n
Sanitization: Evaluation case ID 21 validates that the agent ignores malicious instructions in processed data.\n- [COMMAND_EXECUTION]: The skill requires the execution of a local Node.js script (scripts/gemini_search.js) to interact with the Gemini API. This execution is scoped to the skill's own helper script and is necessary for its stated purpose.\n- [EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch grounded search data from Google's official Gemini API and to retrieve content from external URLs. These operations are essential to the skill's functionality and target well-known/trusted services.

gemini-search