The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill allows the agent to navigate and interact with arbitrary web content, which presents a surface for indirect prompt injection attacks.
Ingestion points: External data is ingested via computer-use browser navigate and computer-use browser snapshot as documented in SKILL.md.
Boundary markers: The skill does not implement specific delimiters or instructions to ignore embedded commands within the processed web content.
Capability inventory: The skill provides significant control over the host system, including the ability to press controls, set field values, click web elements, and write screenshots to the /tmp/ directory.
Sanitization: The documentation notes that snapshots are limited in node depth and include heuristic redaction of sensitive fields, though this does not prevent instruction injection.
[COMMAND_EXECUTION]: The skill's primary function is the execution of a suite of commands via the computer-use CLI to automate user interface interactions on macOS. Examples include computer-use launch, computer-use observe, and computer-use press in SKILL.md.

computer-use