The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied molecular datasets and SMILES strings, which represents an indirect prompt injection surface.
Ingestion points: Molecular datasets and SMILES strings are ingested from user code as described in SKILL.md and the tasks in evals/evals.json.
Boundary markers: The instructions explicitly tell the agent to "Parse SMILES with RDKit and reject failed parses (None)", providing a validation step.
Capability inventory: The skill's operations are restricted to GPU-accelerated molecular calculations (fingerprints, similarity matrices, conformer generation, etc.). No generic file-system write, network exfiltration, or arbitrary command execution capabilities are present.
Sanitization: Inputs are sanitized through the RDKit molecular parsing engine.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing dependencies like nvmolkit, torch, and rdkit using standard package managers.
The installation guidance points to official and vendor-owned resources, including pytorch.org and nvidia-bionemo.github.io.
[OBFUSCATION]: The skill.oms.sig file contains large Base64-encoded strings within the rawBytes fields.
These strings are standard X.509 cryptographic certificates used for verifying the skill's integrity and do not represent malicious code obfuscation.

nvmolkit-usage