The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The file zinc-database_audit_result_v1.json contains numerous self-referential claims asserting the skill's safety and successful security status (e.g., 'security: PASS', 'gate: PASS', 'deployable: true'). These claims target the security auditing process itself and are evaluated as an attempt to influence the analyzer's verdict.
Ingestion points: The audit results file is ingested by the security analyzer; external chemical data is ingested from the ZINC API via pandas.read_csv.
Boundary markers: Absent; the skill does not use delimiters or instructions to ignore embedded content in the processed data.
Capability inventory: The skill uses subprocess.run to execute curl commands and provides examples for wget and aria2c for file downloads. It also uses the curl @ syntax to read local files (zinc_ids.txt).
Sanitization: No explicit sanitization or validation is performed on the data fetched from the ZINC API before processing with pandas or re.
[COMMAND_EXECUTION]: The skill's primary functionality relies on executing shell commands. SKILL.md contains a Python implementation using subprocess.run(['curl', ...]) to fetch data, and references/api_reference.md provides shell command examples for curl, wget, aria2c, gunzip, and obabel (OpenBabel).
[EXTERNAL_DOWNLOADS]: The skill is designed to download large chemical datasets and 3D structure files from https://cartblanche22.docking.org and https://files.docking.org. These are well-known academic resources for cheminformatics, but represent an external data dependency.

zinc-database