The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external websites through tools such as scrape, google_search, and various e-commerce SDE tools. This content is processed by the agent, creating a surface for indirect prompt injection attacks.
Ingestion points: Tools that fetch external web content (e.g., scrape, google_search, amazon_product, walmart_search).
Boundary markers: The instructions recommend structured outputs (JSON, Markdown) and synthesis, which provides some implicit structure, though it does not provide explicit sanitization for malicious instructions embedded in the target pages.
Capability inventory: The skill allows network requests to fetch data and performs shell-based server execution during setup.
Sanitization: The skill lacks explicit instructions for filtering or escaping malicious content within the scraped results.
[EXTERNAL_DOWNLOADS]: The documentation provides standard setup instructions that involve downloading the MCP server implementation from vendor-controlled or well-known registries.
Evidence: Instructions to run npx mcp-remote https://mcp.scraperapi.com/mcp and pip install scraperapi-mcp-server in references/setup.md.
[COMMAND_EXECUTION]: The skill requires starting an MCP server via shell commands during the installation phase.
Evidence: Command-line instructions for starting the remote server via npx and the local server via python -m scraperapi_mcp_server.
[DATA_EXFILTRATION]: The skill documentation describes a feature (callbackUrl in crawler_job_start) that allows scraped data to be sent to a user-defined external endpoint.
Evidence: references/crawler.md explicitly warns the user that this feature transmits page results (including potential PII or sensitive data) over the network and requires user approval.

scraperapi-mcp