The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides installation commands for several Python packages (torch, torch-geometric, torchdrug, rdkit) using pip install. Users should ensure these are installed in a controlled environment and verify package integrity.
[DYNAMIC_EXECUTION]: The 'Pretrained Models and Transfer Learning' section demonstrates the use of torch.load() to load model weights. This function uses Python's pickle module, which can execute arbitrary code during the loading process if the .pth file is malicious or has been tampered with.
[EXTERNAL_DOWNLOADS]: Several examples utilize dataset loaders (e.g., datasets.BBBP, datasets.Davis) that automatically fetch benchmark molecular data from external repositories like MoleculeNet to local directories (e.g., ~/data/).
[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for reading data from external CSV files (pd.read_csv) and molecular SMILES strings. This represents a vulnerability surface where untrusted data could be introduced into the agent's context.
Ingestion points: pd.read_csv("compounds.csv"), pd.read_csv("assay_data.csv"), and SMILES string inputs in data.Molecule.from_smiles().
Boundary markers: No explicit delimiters or instructions to ignore embedded content are present in the provided examples.
Capability inventory: The skill possesses capabilities for file system access (pd.read_csv), command execution (pip install), and dynamic code execution (torch.load).
Sanitization: The examples do not demonstrate sanitization or validation of the input SMILES or CSV content before processing.

torchdrug