dsh-plugin-developer

Pass

Audited by Gen Agent Trust Hub on Nov 7, 2026

Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: Example development scripts contain hardcoded absolute file paths specific to the author's local environment (e.g., C:\Users\49707\AppData\Local\... in assets/examples/weather-plugin/scripts/link_deps.py). While this reveals information about the developer's machine configuration, it is an artifact of the example code and not a malicious data harvesting attempt.
  • [COMMAND_EXECUTION]: The skill instructs the agent to use shell commands for development tasks, such as creating directory junctions via cmd /c mklink and executing external CLI tools like ffmpeg and the ByteDance dreamina-canvas CLI. These operations are intended for the skill's primary purpose of media processing and environment setup.
  • [INDIRECT_PROMPT_INJECTION]: The skill architecture creates an attack surface where plugins process untrusted user data (like myths or stories). This risk is explicitly addressed in the skill's documentation, which provides mitigation strategies.
  • Ingestion points: Plugins ingest untrusted user stories or novels in assets/examples/dsh-seedance-studio/skills/smy-seedance-storyboard/SKILL.md.
  • Boundary markers: The skill recommends using <data> tags and explicit warnings in references/host-integration.md to prevent the agent from obeying embedded instructions.
  • Capability inventory: Plugins generated by this skill possess capabilities for file system modification and subprocess execution (seen in dreamina.js and ffmpeg.js).
  • Sanitization: The documentation provides patterns for validating and normalizing model output before passing it to executable tools.
  • [EXTERNAL_DOWNLOADS]: The skill provides installation commands for third-party CLI tools using piped remote execution (irm | iex) from ByteDance's jianying.com domain, which is a well-known service in the Chinese AI ecosystem.
  • [DYNAMIC_EXECUTION]: Testing utilities in the skill utilize new Function() to execute bundled code in sandboxed testing environments, and dynamic resolution of binary paths for local tool execution.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 7, 2026, 06:45 AM