The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the sp CLI tool to perform system-level operations, including listing apps and windows, and launching processes with arbitrary arguments and environment variables.
[REMOTE_CODE_EXECUTION]: The sp computer-use page execute-javascript command allows for the execution of arbitrary JavaScript within browser contexts, which is a form of dynamic code execution.
[DATA_EXFILTRATION]: The skill can access potentially sensitive data by taking screenshots (sp computer-use snapshot) and extracting text content from browsers (sp computer-use page get-text). This data is processed by the agent and could be used to expose private information found on the user's screen or within web applications.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through the processing of untrusted data from UI elements and web pages.
Ingestion points: Accessibility element properties (role, title, value) in references/snapshot-and-elements.md and page text/DOM content in references/actions.md.
Boundary markers: Absent; there are no instructions to the agent to distinguish between its own system instructions and commands potentially found within the processed UI or web data.
Capability inventory: Extensive system interaction capabilities including shell commands via the sp CLI tool, simulated input (typing, clicking, key presses), and browser JavaScript execution.
Sanitization: No sanitization or validation of external data is specified before it is ingested and used by the agent to make decisions.

supaterm-computer-use