The Agent Skills Directory

[COMMAND_EXECUTION]: The Python integration examples in SKILL.md and references/api_reference.md provide templates that use subprocess.run to execute curl commands. These snippets construct URLs using f-strings with parameters like smiles and zinc_id without implementing sanitization or proper URL encoding. This creates a surface for argument injection where a malicious chemical string containing shell metacharacters or curl flags (e.g., -o) could manipulate the command execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted chemical identifiers which are then interpolated into tool calls.
Ingestion points: smiles, zinc_id, and catitem_id parameters in functions located in SKILL.md and references/api_reference.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the provided templates to prevent the agent from misinterpreting injected content.
Capability inventory: The skill facilitates network requests and shell execution via subprocess.run(['curl', ...]) across all integration scripts.
Sanitization: Input validation and escaping are absent in the primary code templates, although urllib.parse.quote() is briefly mentioned in a troubleshooting section.
[EXTERNAL_DOWNLOADS]: Fetches chemical property data and 3D molecular structures from the University of California, San Francisco (UCSF) official ZINC database domains (cartblanche22.docking.org, files.docking.org).

zinc-database