se-dev-script

Warn

Audited by Gen Agent Trust Hub on Apr 28, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The Prepare.bat script downloads an installation script for the uv tool from https://astral.sh/uv/install.ps1 and a busybox.exe executable from https://frippery.org/files/busybox/busybox64u.exe. While astral.sh is a well-known source for Python tooling, frippery.org is an external domain used to host BusyBox binaries for Windows.
  • [COMMAND_EXECUTION]: The skill performs several system-level operations during its preparation phase.
  • Prepare.bat queries the Windows Registry (HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows\CurrentVersion\Uninstall\Steam App 244850) to locate the game's installation path.
  • It uses mklink /J to create directory junctions for SteamScripts and LocalScripts, connecting the skill folder to the game's Workshop and AppData directories.
  • run_prepare.sh and Prepare.bat execute shell commands to synchronize the Python environment using uv sync and manage file cleanup.
  • The skill provides the agent with the ability to run various UNIX-like commands through busybox.exe and specialized search operations via Python scripts.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted data from the Space Engineers Workshop.
  • Ingestion points: Content is read from Script.cs files within the SteamScripts/ and LocalScripts/ directories by index_scripts.py.
  • Boundary markers: Absent. The search results generated by search_scripts.py output raw code snippets directly into the agent's context without delimiters or instructions to ignore embedded commands.
  • Capability inventory: The skill allows for shell command execution (via BusyBox) and Python script execution, which could be abused if the agent obeys instructions found within the indexed scripts.
  • Sanitization: Absent. The indexing and search processes do not filter or sanitize the source code content for potential malicious instructions or prompt injection patterns.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 28, 2026, 10:19 PM