humanoid-motion
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill uses the
transformerslibrary to fetch thefacebook/dinov2-with-registers-basemodel from Hugging Face. Meta (Facebook) and Hugging Face are recognized as trusted providers for machine learning assets. - [COMMAND_EXECUTION]: The main automation script
twocam.pyusessubprocess.Popento manage a multi-stage pipeline, including joint detection, fitting, and Blender rendering. Command arguments are handled as lists to prevent shell injection vulnerabilities. - [DYNAMIC_EXECUTION]: A
pickle.loadscall is present in the training script's self-test suite. This is used to verify that internal data structures are compatible with Python's multiprocessing requirements for data loading and does not involve processing untrusted external data. - [INDIRECT_PROMPT_INJECTION]: The skill includes an attack surface where user-provided action descriptions are incorporated into templates for video generation. While this is inherent to the tool's purpose, it represents a path for potential instruction injection into the video generation model.
- Ingestion points: User-provided action strings in
make_prompt.pyandfilm.pyused to generate video content. - Boundary markers: The prompt template in
make_prompt.pyuses clear "Action:" delimiters. - Capability inventory: Subprocess execution in
twocam.py, file writes across several scripts for managing the animation pipeline. - Sanitization:
make_prompt.pyperforms basic whitespace trimming and ensures sentence termination. - [SAFE]: The static analysis alert for
eval()indetect.pyis a false positive. The code calls the.eval()method on a PyTorch neural network to disable training-specific layers (like dropout) during inference, which is a standard procedure in machine learning and unrelated to the dangerous Pythoneval()function.
Audit Metadata