The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides instructions to set broad directory permissions (chmod 777) on the local directory used to cache model weights. This allows any user on the host system to read, write, or execute files within that directory.
Evidence (SKILL.md): chmod 777 "${LOCAL_NIM_CACHE}"
Evidence (references/api.md): chmod 777 "${LOCAL_NIM_CACHE}"
[INDIRECT_PROMPT_INJECTION]: The skill processes molecular notation (SMILES and SAFE strings) returned from an external API. This presents a potential attack surface where malicious data could influence the agent's behavior during data aggregation or reporting tasks.
Ingestion points: API response data parsed in the result processing loop (SKILL.md).
Boundary markers: None present.
Capability inventory: Includes shell command execution (Bash) and file writing (Write).
Sanitization: No explicit validation or sanitization is performed on the chemical strings before they are printed to the console or written to files.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill downloads container images and approximately 20 GB of model weights from the NVIDIA Container Registry (nvcr.io). As these resources originate from the vendor's official infrastructure, this is a standard and expected operational procedure for the GenMol service.

genmol-nim