The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation link https://jeffallan.github.io/claude-skills/skills/data-ml/ml-pipeline/ found in SKILL.md is flagged as a malicious URL by automated reputation scanners (URL:Blacklist). Additionally, the file SKILL.md itself has been identified as potentially malicious (FileRepMalware) by security software.
[DYNAMIC_EXECUTION]: The skill's templates in SKILL.md, references/feature-engineering.md, references/pipeline-orchestration.md, and references/training-pipelines.md utilize the pickle and joblib libraries for model and pipeline serialization. The implementation of pickle.load() and joblib.load() on file paths that can be influenced by untrusted external data (such as ML model artifacts or feature store paths) creates a significant risk of arbitrary code execution.
[INDIRECT_PROMPT_INJECTION]: The skill provides templates for data ingestion and quality validation using Great Expectations (in SKILL.md and references/feature-engineering.md) and custom validation logic (in references/model-validation.md). These ingestion points lack explicit boundary markers or sanitization to prevent adversarial instructions embedded in training datasets or model metadata from influencing the agent's context or downstream logic.
[COMMAND_EXECUTION]: Orchestration templates for Apache Airflow and Kubeflow in references/pipeline-orchestration.md perform file system operations, such as shutil.copy to sensitive directories like /models/production/. If input arguments such as run_id or model_name are not strictly sanitized, an attacker could potentially achieve unauthorized file write capabilities.

ml-pipeline