The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on external command execution to perform its primary function. It specifically instructs the agent to use exec: peekaboo screenshot (a macOS tool) and other capture tools like browser: screenshot and nodes: screen_record to obtain visual data from the environment.
[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves capturing untrusted visual data from external websites or application interfaces and passing it directly to a vision model for analysis. This creates a vulnerability where an attacker could embed malicious text or instructions within a UI to influence the agent's review or subsequent actions.
Ingestion points: Untrusted visual content is ingested via browser: screenshot, nodes: screen_record, and exec: peekaboo screenshot as described in SKILL.md.
Boundary markers: No boundary markers or specialized delimiters are used to separate the vision model's instructions from the content of the image being analyzed.
Capability inventory: The skill has the capability to capture the user's screen, record activity, and read files via the image tool (SKILL.md).
Sanitization: There is no evidence of sanitization, filtering, or validation performed on the captured visual data before it is sent to the vision model.

visual-qa