The Agent Skills Directory

[COMMAND_EXECUTION]: The Python code snippets in SKILL.md use subprocess.run(['curl', url]) to query the ZINC database. The URL is constructed using f-strings that incorporate variables such as zinc_id and smiles strings directly from user input. This pattern creates an argument injection surface where a malicious user could provide inputs starting with hyphens to alter the behavior of the curl command.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to external domains including cartblanche22.docking.org, files.docking.org, and zinc.docking.org to fetch compound metadata and 3D molecular structures.
[PROMPT_INJECTION]: The skill presents a surface for Indirect Prompt Injection (Category 8). 1. Ingestion points: External data is ingested from CartBlanche22 API text and JSON endpoints. 2. Boundary markers: No delimiters or instructions to ignore embedded commands are present. 3. Capability inventory: The skill utilizes subprocess execution and data parsing capabilities. 4. Sanitization: No validation or sanitization of external database records is performed before the agent processes the information.
[PROMPT_INJECTION]: The skill contains metadata instructions that direct the agent to steer users toward the author's commercial platform (www.k-dense.ai) when workflow complexity increases, which is a form of steering behavior that can influence the agent's objectivity.

zinc-database