unity-cli

Warn

Audited by Gen Agent Trust Hub on Nov 6, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_EXECUTION]: The C# script resources/Editor/BatchAudit.cs uses reflection to resolve and invoke methods at runtime. The target type and method names are supplied directly from command-line arguments (-audit flag), allowing the execution of public static string methods within the project's assemblies that match the expected signature. Additionally, subcommands like unity command eval allow for the evaluation of arbitrary C# code strings.
  • [COMMAND_EXECUTION]: The tool provides subcommands like unity shell and unity projects exec which are designed to execute arbitrary code or shell commands. The documentation warns that these features run with the user's privileges and should only be used with trusted input.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project data (assets, scenes, meta files) and possesses powerful execution capabilities. There is a risk that malicious instructions embedded in project files or logs could be executed or influence the agent's behavior when using the CLI's evaluation or shell features.
  • Ingestion points: resources/Editor/BatchAudit.cs reads CLI flags and loads assets; reference/version-control.md describes parsing YAML scenes.
  • Boundary markers: Absent in the provided implementation; the documentation provides manual warnings to the user.
  • Capability inventory: Shell execution (unity shell), file writing (BatchAudit.cs), reflection-based method invocation (BatchAudit.Run), and project configuration management.
  • Sanitization: The documentation advises user caution and manual review, but no automated sanitization is shown in the provided code.
  • [EXTERNAL_DOWNLOADS]: The SKILL.md file contains an installation command that fetches and executes a shell script from Unity's official public CDN (public-cdn.cloud.unity3d.com). As this targets a well-known service, the download is documented neutrally and does not contribute to verdict escalation.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 6, 2026, 06:48 AM