The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a replayability feature that encourages the execution of shell commands stored within a generated JSON manifest file (run.json). The usage of $(jq -r .cmd_replay $RUN_DIR/run.json) allows for the execution of dynamically generated strings, which could be exploited if the manifest content is manipulated by a malicious actor.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted molecular data (SMILES) provided by the user, which is then passed to internal featurization and embedding scripts.
Ingestion points: SMILES data is read from the CSV file path provided via the --csv argument in SKILL.md.
Boundary markers: No explicit delimiters or isolation instructions are mentioned for the interpolation of external molecular strings into the workflow.
Capability inventory: The skill has significant execution capabilities, including running Python scripts and system commands inside a Docker container via the kermt_container.sh utility.
Sanitization: The workflow includes a validation step using RDKit to check SMILES strings, producing a validity.npy file to flag parsing errors, which serves as a partial mitigation.
[EXTERNAL_DOWNLOADS]: The skill optionally downloads a pre-trained model bundle from the official repository nvidia/NV-KERMT-70M-v2 on Hugging Face. This download is conducted after obtaining explicit user consent or through an explicit opt-in flag, and it targets a well-known, trusted organization.

kermt-embed