The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content retrieved from the web via the Tavily API, which creates a potential surface for indirect prompt injection if retrieved websites contain malicious instructions.
Ingestion points: Data is ingested through the tavily_search, tavily_extract, tavily_crawl, and tavily_research tools (implemented in their respective scripts in the scripts/ directory).
Boundary markers: The scripts use markdown headers (e.g., ## Sources, ### [URL]) to delimit results, but they do not explicitly instruct the agent to ignore embedded instructions within the content.
Capability inventory: The ingested data is printed to the agent's context. The skill scripts do not perform file writes, command execution, or additional network operations based on the retrieved content.
Sanitization: scripts/crawl.mjs and scripts/search.mjs truncate long content blocks, which limits the volume of potentially malicious data, though no specific filtering of natural language instructions is applied.

tavily-search