The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/adme_predictor.py uses pickle.load() to deserialize machine learning models from the models/ directory. The pickle module is notoriously insecure because it can execute arbitrary code embedded in the data stream during deserialization. If an attacker manages to modify the local .pkl files distributed with the skill, they could achieve arbitrary code execution when the agent runs the prediction tool.
[EXTERNAL_DOWNLOADS]: In scripts/train_tdc_models.py, the skill uses the PyTDC library to fetch molecular datasets from the Therapeutics Data Commons (TDC) hosted at tdcommons.ai. This is a standard procedure for retrieving scientific data for training the predictive models and is considered a legitimate operation within the context of the skill's functionality.

adme-prediction