avalonia

Warn

Audited by Gen Agent Trust Hub on Jul 22, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The preview-server sub-skill includes Python scripts (scaffold_host.py, render_variants.py) that utilize subprocess.run to execute dotnet CLI commands for project creation, package management, and application execution.
  • [REMOTE_CODE_EXECUTION]: The render_variants.py script identifies and installs NuGet packages from a packages.txt file generated by the agent at runtime. This allows for the dynamic inclusion and execution of arbitrary third-party libraries from the NuGet registry.
  • [DYNAMIC_EXECUTION]: The skill facilitates the generation of C# and XAML source code into a local project structure (.preview/AvaloniaPreviewHost/) which is then compiled and executed using dotnet run to generate visual previews. This 'code generation and execution' loop is a known vector for code injection if input is not strictly validated.
  • [EXTERNAL_DOWNLOADS]: The documentation and scripts encourage the resolution and download of numerous community-provided Avalonia UI packages from external repositories (NuGet). While many are well-known (e.g., SukiUI, FluentAvalonia), the system allows for the resolution of any package name provided in the variant configuration.
  • [INDIRECT_PROMPT_INJECTION]: The 'Preview Server' feature creates an attack surface for indirect prompt injection. A malicious instruction could potentially influence the agent to generate a UI variant containing malicious package references or XAML-based exploits, which are then processed and executed by the provided utility scripts.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 22, 2026, 05:47 PM
Security Audit — agent-trust-hub — avalonia