torchdrug

Pass

Audited by Gen Agent Trust Hub on Jul 10, 2026

Risk Level: SAFEREMOTE_CODE_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill demonstrates the use of torch.load() to load model weights. Since torch.load() uses Python's pickle module for deserialization, it can execute arbitrary code if the loaded file is malicious.
  • [PROMPT_INJECTION]: The skill includes code for processing molecular data from external CSV files, creating an indirect prompt injection surface. (1) Ingestion points: pd.read_csv('compounds.csv') and CustomDataset loading SMILES from CSV. (2) Boundary markers: None present. (3) Capability inventory: File system writing (checkpoint saving) and command execution via package installation. (4) Sanitization: None explicitly shown for the CSV input.
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions to download official scientific datasets (MoleculeNet) and install packages from well-known registries. These are standard practices for the intended use case.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 10, 2026, 03:36 PM
Security Audit — agent-trust-hub — torchdrug