lsfg-android-frame-generation

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PRIVILEGE_ESCALATION]: The skill instructs the agent to set up and grant several high-privilege Android permissions:
  • SYSTEM_ALERT_WINDOW (Display over other apps) and TYPE_ACCESSIBILITY_OVERLAY allow the application to draw content over the entire screen.
  • LsfgAccessibilityService uses the Android Accessibility Service framework, which can potentially intercept user interactions.
  • Integration with Shizuku is used to access restricted system APIs for frame timing diagnostics, which requires a privileged side-channel.
  • [COMMAND_EXECUTION]: The skill documentation includes shell commands for project setup and deployment:
  • git clone --recurse-submodules for repository retrieval.
  • ./gradlew for building the Android application and native components.
  • adb install for pushing the compiled APK to a connected device.
  • [EXTERNAL_DOWNLOADS]: The project fetches its source code and required submodules from a public repository on GitHub (https://github.com/FrankBarretta/LSFG-Android.git).
  • [DYNAMIC_EXECUTION]: The application employs native library loading and runtime resource processing:
  • System.loadLibrary("lsfg_android") is used to load the JNI bridge.
  • The application processes a user-provided Lossless.dll via the Storage Access Framework to extract shaders into private storage at runtime.
  • [INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface for indirect injection via processed data:
  • Ingestion points: The application captures live screen data via MediaProjection and ingests an external Lossless.dll file.
  • Boundary markers: None are present in the processing logic described.
  • Capability inventory: The skill has access to Vulkan GPU contexts, native code execution (JNI), and the ability to draw over the system UI.
  • Sanitization: The skill mentions shader extraction but does not detail sanitization or integrity checks for the user-provided binary file.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 10:36 AM