tao-run-automl
Warn
Audited by Gen Agent Trust Hub on Aug 25, 2026
Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
- [DYNAMIC_EXECUTION]: The skill generates and executes Python runner scripts at runtime based on user configurations and model metadata. Evidence: SKILL.md and references/automl-intent-algorithms.md describe the creation of unique timestamped runner scripts like run_{RUN_NAME}_{TIMESTAMP}.py.
- [UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill recommends installing external Python packages to enable AutoML and LLM functionality. Evidence: SKILL.md and references/automl-preflight-concepts.md provide shell commands for installing packages such as nvidia-tao-automl, openai, and wandb using a local version resolution script.
- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill interacts with sensitive environment variables for external service authentication. Evidence: references/automl-runner-configuration.md and references/automl-intent-algorithms.md mention the use of AUTOML_LLM_API_KEY, NVIDIA_API_KEY, WANDB_API_KEY, and HF_TOKEN. Instructions are provided to verify these variables without printing their contents to the console.
- [INDIRECT_PROMPT_INJECTION]: The skill processes data from external LLM endpoints and logs for its optimization algorithms, which could be influenced by external data. Evidence: references/automl-advanced-monitoring.md describes components like KnowledgeRetriever and SpecPrescreener that use LLM outputs. 1. Ingestion points: KnowledgeRetriever and SpecPrescreener (references/automl-advanced-monitoring.md). 2. Boundary markers: Absent. 3. Capability inventory: Read, Bash, and Write tool access for script generation and subprocess execution (SKILL.md). 4. Sanitization: Filename slugification (references/automl-intent-algorithms.md).
Audit Metadata