The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation link in SKILL.md (https://jeffallan.github.io/claude-skills/skills/data-ml/ml-pipeline/) points to a domain currently blacklisted by security scanners (URLite). Additionally, the file reputation scanner flagged SKILL.md itself as potentially malicious content.
[DYNAMIC_EXECUTION]: The FeaturePipeline.load method in references/feature-engineering.md uses pickle.load() to deserialize pipeline objects from a file path provided at runtime. Using pickle to load untrusted data is a critical vulnerability that can result in arbitrary code execution.
[INDIRECT_PROMPT_INJECTION]: The skill implements various data ingestion points using pandas.read_csv() and pandas.read_parquet() in SKILL.md and references/pipeline-orchestration.md without any sanitization or boundary markers. This exposes the agent to adversarial data that could be used to manipulate the pipeline construction or subsequent analysis.
Ingestion points: SKILL.md, references/pipeline-orchestration.md.
Boundary markers: None present in code templates.
Capability inventory: pickle.load (Dynamic Execution), shell command execution via Airflow and Kubeflow DAG definitions.
Sanitization: None present.
[REMOTE_CODE_EXECUTION]: Kubeflow and Prefect pipeline templates in SKILL.md and references/pipeline-orchestration.md are configured to download and install arbitrary Python packages from external registries at runtime via the packages_to_install parameter, introducing supply chain risks if the registry or package names are targeted.
[COMMAND_EXECUTION]: The orchestration templates for Airflow and Kubeflow facilitate the execution of complex Python logic and shell commands based on pipeline parameters, which could be exploited if malicious inputs flow into the orchestration layer.

ml-pipeline