The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the UI of Windows applications, which presents a surface for indirect prompt injection if an application contains text designed to manipulate the agent's logic.
Ingestion points: SKILL.md defines tools such as windows_snapshot, windows_find, windows_get_text, windows_get_table, and windows_clipboard that ingest data from the user's active desktop environment.
Boundary markers: Absent. The instructions do not specify the use of delimiters or warnings for the agent to ignore instructions potentially embedded in the UI content.
Capability inventory: The toolset includes process execution (windows_launch), keyboard and mouse simulation (windows_click, windows_type, windows_send_keys), and system data access (windows_screenshot, windows_clipboard).
Sanitization: Absent. There are no directives to sanitize or validate retrieved text before processing.
[SAFE]: The skill demonstrates high security awareness by implementing a check for committing actions (e.g., 'Delete', 'Pay', 'Confirm') and requiring user confirmation for side effects. It also follows the best practice of cleaning up the environment using windows_end and discarding temporary changes made during exploration.

exploring-windows