agent-desktop
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The skill includes commands that allow the agent to read potentially sensitive information from the host system, such as the system clipboard, desktop screenshots, and notification content. While no automatic external transmission is defined, the exposure to the agent is significant.
- Evidence:
commands-system.mddefinesclipboard-getandlist-notifications.commands-observation.mddefines thescreenshotcommand. - [INDIRECT_PROMPT_INJECTION]: The agent is instructed to process data from the desktop environment, which may contain untrusted content from third-party applications or websites that could attempt to influence the agent's behavior.
- Ingestion points: The agent reads data via
snapshot,find, andgetincommands-observation.md, and throughlist-notificationsandclipboard-getincommands-system.md. - Boundary markers: There are no explicit instructions for the agent to use delimiters or ignore instructions embedded within the UI text it reads.
- Capability inventory: The agent can execute powerful actions such as
launch,click,type,press, andclose-appbased on its observations. - Sanitization: No sanitization or validation of the ingested UI text is mentioned before it is used to determine the agent's next action.
- [COMMAND_EXECUTION]: The skill leverages a CLI tool to perform high-privilege actions on the desktop, including application lifecycle management and simulation of mouse and keyboard events.
- Evidence:
commands-system.mddetails commands forlaunch,close-app, and window management.commands-interaction.mddetails input simulation commands likeclick,type, andpress.
Audit Metadata