The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes search results from the public web, which are untrusted third-party inputs that may contain malicious instructions.
Ingestion points: Untrusted data enters the agent context through search result titles, URLs, and metadata retrieved via curl calls in SKILL.md and EXAMPLES.md.
Boundary markers: The skill includes explicit "Safety Notes" in SKILL.md, EXAMPLES.md, and REFERENCE.md warning the agent to treat results as untrusted remote input.
Capability inventory: The skill uses the Bash tool to perform network requests, parse JSON with jq, and manipulate strings with sed.
Sanitization: While the skill uses jq for structured parsing, the provided examples demonstrate shell interpolation of search results into commands and file system paths, relying on the LLM's adherence to the safety instructions for security.
[COMMAND_EXECUTION]: The skill uses the Bash tool to interact with the Search API and perform post-processing of results.
Evidence: Multiple examples in SKILL.md and EXAMPLES.md utilize curl, jq, shell pipes, and redirection to local files to manage the search workflow.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to external endpoints to facilitate searching and fetching content.
Evidence: The skill performs POST requests to https://api.browserbase.com/v1/search and https://api.browserbase.com/v1/fetch. These destinations correspond to the skill author's official infrastructure.

search