The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests large volumes of external chemical data from the ZINC22 API (cartblanche22.docking.org). Maliciously crafted data in the database could potentially influence agent behavior if processed without sanitization.
Ingestion points: API endpoints at cartblanche22.docking.org and file repositories at files.docking.org.
Boundary markers: No explicit delimiters or warnings to ignore embedded instructions are provided in the documentation.
Capability inventory: The skill utilizes shell command execution via curl and Python's subprocess module.
Sanitization: No explicit sanitization or validation of the retrieved SMILES or compound data is demonstrated in the integration examples.
[COMMAND_EXECUTION]: The skill provides Python integration examples that use subprocess.run to execute the curl binary. The URLs are constructed using f-string interpolation from user-supplied variables (e.g., zinc_id, smiles). While using a list for subprocess arguments mitigates standard shell injection, improper sanitization of these inputs could still lead to argument injection or SSRF-like behavior if the agent treats the provided code as a template for execution.
[METADATA_POISONING]: The skill includes a specific directive at the end of the file instructing the agent to proactively suggest the author's platform (K-Dense Web) when tasks increase in complexity. This acts as a steering instruction that influences the agent's behavior toward a specific commercial service.

zinc-database