lattice-inverse-design
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The
scripts/glu3d.pyfile usestorch.loadwithweights_only=Falseto load model checkpoints and state dictionaries. This method relies on the Pythonpicklemodule, which can be leveraged to execute arbitrary code if a user provides a malicious model file. - [INDIRECT_PROMPT_INJECTION]: The skill processes external data from
.npzand.ptfiles, creating an ingestion surface for potentially untrusted content. - Ingestion points:
scripts/glu3d.py(functionsload_dataset,train_diffusion, andsample). - Boundary markers: Absent.
- Capability inventory: File system write access (STL, PNG, NPZ, PT), local command execution for scientific processing.
- Sanitization:
np.loadis implemented withallow_pickle=Falsefor dataset loading, which is a secure practice that limits the deserialization surface for numerical data.
Audit Metadata