os-integrations

Warn

Audited by Gen Agent Trust Hub on Nov 2, 2026

Risk Level: MEDIUMPERSISTENCEDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PERSISTENCE]: The skill provides C# implementation details in recipes.md for a 'Launch at login' feature. This involves writing the application's executable path to the Windows Registry at HKCU\Software\Microsoft\Windows\CurrentVersion\Run, allowing the application to persist across reboots.
  • [DATA_EXFILTRATION]: Functional code is provided for low-level system hooks. Specifically, recipes.md and overlays.md demonstrate the use of SetWindowsHookEx(WH_KEYBOARD_LL) and SetWindowsHookEx(WH_MOUSE_LL). These hooks intercept all keystrokes and mouse movements globally, which can be used to monitor user activity or function as a keylogger in non-elevated windows.
  • [DATA_EXFILTRATION]: The capture.md and recipes.md files provide detailed recipes for high-fidelity screen capture and recording. This includes using GDI BitBlt to bypass certain capture borders and Windows.Graphics.Capture to record screen activity to MP4 files via the GPU.
  • [COMMAND_EXECUTION]: The recipes.md file contains a 'Debug bridge' implementation using NamedPipeServerStream. This creates a listener that executes commands received over a named pipe on the UI thread, which establishes a primitive for local command injection.
  • [INDIRECT_PROMPT_INJECTION]: The skill facilitates the registration of custom URL protocols (e.g., myapp://) in recipes.md. This creates an attack surface where maliciously crafted links can trigger actions within the application via external sources.
  • Ingestion points: Registration of UrlScheme (recipes.md), DebugBridge named pipe (recipes.md), and ClipboardWatcher (recipes.md).
  • Boundary markers: The documentation advises treating input as 'untrusted,' but the provided code examples do not include explicit boundary enforcement or safety delimiters.
  • Capability inventory: The skill includes code for SendInput, SetWindowsHookEx, BitBlt, registry manipulation, and named pipe communication.
  • Sanitization: While basic string replacement is shown for SQL queries in system.md, the general input logic for command handling in the DebugBridge lacks robust sanitization against potentially malicious input strings.
  • [EXTERNAL_DOWNLOADS]: Documentation in system.md and overlays.md references the inclusion of NuGet packages like System.Data.OleDb and System.Drawing.Common to support OS integration features.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 2, 2026, 04:41 AM