hinge-liker
Pass
Audited by Gen Agent Trust Hub on Oct 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The script uses subprocess.run to execute commands via the Android Debug Bridge (adb) and curl. The path to the adb binary is user-configurable through the ADB_PATH environment variable or the --adb CLI argument, allowing the script to call a user-specified executable.
- [CREDENTIALS_UNSAFE]: SKILL.md advises users to hardcode the GEMINI_API_KEY directly into cron payloads when scheduling automated sessions, which results in sensitive API credentials being stored in plain text within system configuration files.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted dating profile content (text and images) from Hinge using an AI model, creating a surface for indirect prompt injection. Ingestion points: Screenshots and UI XML dumps are pulled from the emulator via adb. Boundary markers: The AI prompt specifies a JSON response format. Capability inventory: The script executes shell commands (adb) and network requests (curl). Sanitization: The AI-generated comment is escaped for quotes and spaces before being passed to the adb shell input command.
- [EXTERNAL_DOWNLOADS]: The skill sends profile data to the Google Gemini API (generativelanguage.googleapis.com) for analysis. Interaction with this well-known service is used for the skill's primary vision-based profile evaluation functionality.
Audit Metadata