The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documentation and examples include common patterns for loading data from external files, such as CSV files via pandas.read_csv. This introduces an inherent surface for indirect prompt injection if the agent is instructed to analyze data from untrusted or attacker-controlled sources.
Ingestion points: Data loading examples in SKILL.md and references/quick_reference.md.
Boundary markers: None present in the simplified examples.
Capability inventory: Local file system writes for plots (plt.savefig) in scripts/clustering_analysis.py and model training compute.
Sanitization: No sanitization of input data is demonstrated in the basic examples.
[DYNAMIC_EXECUTION]: The reference guide in references/model_evaluation.md documents the use of pickle and joblib for model persistence. While standard in the machine learning ecosystem, loading untrusted serialized files can lead to arbitrary code execution; however, this is presented as a standard documentation feature and not used maliciously.
[METADATA_POISONING]: There is a minor inconsistency between the author listed in the metadata ('K-Dense Inc.') and the expected author context, though this appears to be a documentation discrepancy rather than a malicious attempt to deceive.

scikit-learn