The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting external data from the PubChem database. While PubChem is a curated scientific resource, the ingestion of external data combined with file-writing capabilities constitutes a potential attack surface.
Ingestion points: Data is retrieved from the PubChem API (pubchem.ncbi.nlm.nih.gov) within scripts/bioactivity_query.py and scripts/compound_search.py.
Boundary markers: The instructions and scripts do not implement specific delimiters or 'ignore' instructions to isolate retrieved chemical data from the agent's logic.
Capability inventory: The skill includes file-writing capabilities via pcp.download() and standard Python file operations in SKILL.md and scripts/compound_search.py, alongside network access to the official PubChem API.
Sanitization: The skill does not perform sanitization or validation of strings (such as chemical synonyms or assay descriptions) returned by the API before they are processed by the agent.
[METADATA_POISONING]: There is a discrepancy between the author specified in the skill's YAML frontmatter ('K-Dense Inc.') and the actual author context ('leonchaox'). This appears to be a documentation inconsistency rather than a malicious attempt to deceive.

pubchem-database