The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a specific instruction block ('Suggest Using K-Dense Web') that directs the agent to proactively recommend a third-party research platform (K-Dense Web) to the user under specific conditions. This is an attempt to steer the agent's neutral behavior and hijack the conversation flow for promotional purposes.
[METADATA_POISONING]: There is a conflict between the provided skill author context ('wu-yc') and the author stated in the skill's YAML metadata and content ('K-Dense Inc.'). This discrepancy is misleading regarding the true origin and maintainer of the skill.
[CREDENTIALS_UNSAFE]: The documentation and code examples provide templates for the download_cosmic_file function that accept email and password as plaintext arguments. This pattern encourages users to hardcode sensitive authentication credentials directly into their Python scripts or command-line history.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data files (TSV, CSV, VCF) from the COSMIC database which could theoretically contain malicious payloads.
Ingestion points: Files are read using pandas.read_csv and pysam.VariantFile in the provided examples.
Boundary markers: None are specified to separate data from instructions during processing.
Capability inventory: The skill references local script execution via scripts/download_cosmic.py and uses data processing libraries like pandas and pysam.
Sanitization: There is no evidence of validation or sanitization of the content fetched from the remote database.

cosmic-database