The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill contains multiple instructions to download executable files and installers from an external GitHub repository (github.com/IRNova). Evidence includes links to APK, EXE, DMG, and AppImage files.
[COMMAND_EXECUTION]: The installation instructions require executing shell commands to install or run the downloaded software, such as adb install for Android, running .exe installers on Windows, using hdiutil attach on macOS, and changing permissions with chmod +x on Linux.
[REMOTE_CODE_EXECUTION]: By directing the agent to download and immediately execute binaries from a repository outside of the vendor's own verified infrastructure or the trusted vendors list, the skill establishes a pattern for remote code execution.
[INDIRECT_PROMPT_INJECTION]: The skill includes logic to fetch content from remote URLs and process it as configuration data, creating a vulnerability surface where malicious external content could influence the agent's behavior.
Ingestion points: The SubscriptionManager.updateSubscription method fetches data from a user-provided profile.url, and RadarScanner._fetchIPsFromSources fetches IP lists from a list of external URLs.
Boundary markers: No delimiters or safety instructions (e.g., "ignore embedded commands") are used when processing the downloaded content.
Capability inventory: The skill possesses capabilities to write files (_writeSingBoxConfig), manage system processes (_startSingBox, _stopSingBox), and perform network operations.
Sanitization: While basic regex validation exists for IP addresses in the scanner, the subscription data is Base64 decoded and split without robust structural validation or sanitization before being used to populate internal objects.

nova-client-proxy