The Agent Skills Directory

[COMMAND_EXECUTION]: The Python script examples/basic_example.py uses eval() to process model output from the EnsemblePipeline. This allows for arbitrary code execution if the string returned by the model is compromised.
[COMMAND_EXECUTION]: The documentation file references/sider_categories.md contains a code example that utilizes eval() to parse side-effect probabilities, encouraging insecure implementation patterns.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing untrusted SMILES strings which are used as input for models whose output is directly evaluated. 1. Ingestion points: SMILES strings in predict_admet (examples/basic_example.py). 2. Boundary markers: None present. 3. Capability inventory: Use of eval() in examples/basic_example.py. 4. Sanitization: Model outputs are not sanitized or validated before evaluation.

admet-prediction