The Agent Skills Directory

[DYNAMIC_EXECUTION]: The ovuiinspect module implements an /execute endpoint that utilizes the compile() and exec() functions to run arbitrary Python code provided via HTTP POST requests.
Evidence: The _advance_execute function in ovuiinspect/__init__.py compiles and executes the code payload within a namespace containing the application instance.
Mitigation: The functionality is gated by the OVUIINSPECT_ENABLE_EXECUTE environment variable and binds to the local loopback interface by default.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates a workflow where the agent captures screenshots of an application and performs actions based on the visual content. This creates a vulnerability to indirect prompt injection if the target application displays malicious text or UI elements designed to manipulate the agent's behavior.
Ingestion points: Application state is ingested via the /screenshot and /capture endpoints in ovuiinspect/__init__.py.
Boundary markers: There are no explicit instructions or delimiters in the skill to warn the agent about ignoring potential instructions found within the captured UI.
Capability inventory: The skill provides capabilities for arbitrary Python execution (/execute), system-level UI input injection (/mouse, /keyboard), and application termination (/shutdown).
Sanitization: No validation or sanitization is performed on the strings passed to type or execute commands.
[COMMAND_EXECUTION]: The skill provides tools for low-level input injection including mouse clicks, drags, and keyboard events. While intended for automation, these tools allow the agent to perform any action a user could, including interacting with terminal emulators or settings if they are visible within the ovui application frame.

omniverse-ui-inspector