blender-ai-replication
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The orchestrator script (
scripts/dream_loop.py) usessubprocess.runto execute a local Blender bridge CLI tool. While functional for the skill's purpose, it involves executing shell commands with user-provided paths.\n- [EXTERNAL_DOWNLOADS]: The documentation suggests downloading third-party plugins from GitHub (e.g.,ahujasid/blender-mcp). Users should verify community resources before installation.\n- [INDIRECT_PROMPT_INJECTION]: The modeling loop uses feedback from Vision Language Models (VLMs) to generate the next iteration's prompt, creating a surface for indirect instructions.\n - Ingestion points: VLM critic responses in
scripts/dream_loop.py.\n - Boundary markers: No explicit delimiters isolate the AI feedback from the generated Blender commands.\n
- Capability inventory: The skill can execute scene modifications and file exports via the bridge CLI.\n
- Sanitization: No explicit sanitization of the VLM-provided feedback is implemented before it drives the next round.
Audit Metadata