The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface by ingesting untrusted web content and images via tools like web_extract and vision_analyze. However, it incorporates a safety boundary by explicitly instructing the agent to treat all page text and pixels as untrusted content and never as instructions.\n
Ingestion points: Web extraction (web_extract) and visual analysis (vision_analyze).\n
Boundary markers: Section 7 contains a clear directive to ignore instructions embedded in captured data.\n
Capability inventory: The skill utilizes terminal, web_extract, and vision_analyze to perform its tasks.\n
Sanitization: Recommends visual inspection by the user and narrow selector framing to limit data exposure.\n- [PERSISTENCE]: The skill guides the agent to configure MCP (Model Context Protocol) servers, which involves writing persistent configuration to the environment to enable tool reuse across sessions.\n
Evidence: Section 4 provides a procedure for adding server entries with absolute command paths to the host's configuration harness.\n- [COMMAND_EXECUTION]: The skill requires the terminal tool to install and verify visual verification binaries.\n
Evidence: The instructions guide the agent to perform binary verification, CLI captures, and integrity checks on installed artifacts, emphasizing the use of absolute paths and pre-installation audits to ensure system safety.

agent-visual-verification