The Agent Skills Directory

[SAFE]: The skill describes the use of qt-ai-dev-tools, a specialized command-line utility for UI automation. All identified functionalities (tree inspection, clicking, typing, and screenshotting) are standard for UI testing and developer environments.
[COMMAND_EXECUTION]: Mentions the use of vm run for executing arbitrary system commands such as pytest, systemctl, and app launchers within a virtual machine. This is a legitimate feature for managing a test environment.
[DATA_EXFILTRATION]: Provides instructions for capturing UI screenshots (screenshot -o /tmp/before.png) and widget tree snapshots (snapshot save before) to the local filesystem for debugging and verification. No network exfiltration or sensitive file access was observed.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it reads text content from third-party application UIs, which could theoretically contain malicious instructions.
Ingestion points: Widget names, labels, and text properties are retrieved from applications via the tree, find, text, and state commands.
Boundary markers: None are defined in the instructions for isolating UI content.
Capability inventory: The agent can perform UI interactions (click, type, key) and execute shell commands (vm run).
Sanitization: No sanitization or filtering of the text data ingested via the AT-SPI accessibility bus is described.

qt-app-interaction