codonfm-embed
Pass
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill facilitates access to model checkpoints and source code from established vendor-controlled repositories.\n
- Model weights are referenced from the official vendor organization on Hugging Face (e.g., nvidia/NV-CodonFM-Encodon-1B-v1).\n
- The execution runner is sourced from the vendor's organization on GitHub (NVIDIA-Digital-Bio).\n- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied CSV files containing genetic sequences, creating a surface for data-driven interaction.\n
- Ingestion points: Input data is read from sequence CSV files via the validation script and model runner.\n
- Boundary markers: The skill relies on structured CSV parsing and explicit column requirements (id, ref_seq, value, split).\n
- Capability inventory: The agent prepares and executes shell commands using the model runner.\n
- Sanitization: The validation script enforces strict character-level checks on sequence data, ensuring only valid nucleic acid bases (A, C, G, T, U) are processed, which significantly mitigates the risk of injection.\n- [DYNAMIC_EXECUTION]: The provided test suite dynamically loads and executes the skill's validation script.\n
- Evidence: The unittest suite in tests/test_validate_inputs.py uses importlib to load and exercise the validate_inputs.py script for verification.\n- [COMMAND_EXECUTION]: The skill prepares shell commands for the agent to validate data and run model inference.\n
- Evidence: SKILL.md provides instructions and examples for executing Python-based validation and inference tools locally.
Audit Metadata