The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from Magnite.com which could potentially contain malicious instructions intended to influence the agent's behavior.
Ingestion points: The magnite_com__read_page and magnite_com__get_root tools described in SKILL.md fetch and display content from external web pages.
Boundary markers: The instructions do not provide explicit delimiters or "ignore embedded instructions" warnings to help the agent distinguish between data and instructions.
Capability inventory: The skill provides multiple tools for searching, reading, and interacting with the target site, which could be misused if the agent obeys instructions found in page content.
Sanitization: No evidence of sanitization, filtering, or validation of the fetched external content is present in the skill instructions.
[EXTERNAL_DOWNLOADS]: The skill utilizes remote resources and installation scripts hosted by the vendor.
Evidence: The README.md and SKILL.md files instruct users to install the skill via npx using the vendor's domain (https://unbrowse.ai).
Evidence: The server.json configuration defines a remote MCP server located at https://unbrowse.ai/mcp/magnite.com for tool execution.

magnite