nano-world-model
Fail
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The installation instructions include a
curlcommand to download a binary model file (i3d_torchscript.pt) from a Dropbox URL (https://www.dropbox.com/scl/fi/c5nfs6c422nlpj880jbmh/i3d_torchscript.pt...). Downloading binary executable content from unauthenticated file-sharing sites is a significant security risk, as the source and integrity of the file cannot be verified. - [DYNAMIC_EXECUTION]: The skill provides Python examples using
torch.load()to load model checkpoints (e.g.,checkpoint = torch.load("/path/to/checkpoint.ckpt", ...)). By default,torch.loaduses thepicklemodule, which is vulnerable to arbitrary code execution if the loaded file is malicious. The risk is heightened because the skill suggests downloading checkpoints from an unverified HuggingFace repository (knightnemo/nanowm-b2-dino-wm-pusht-100k). - [COMMAND_EXECUTION]: The skill instructs the agent to clone a remote Git repository (
simchowitzlabpublic/nano-world-model) and execute local Python scripts usingpython src/main.py. This gives the agent the capability to execute the codebase without a formal security review of the repository content.
Recommendations
- AI detected serious security threats
Audit Metadata