tao-run-automl
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes data through training logs and experiment history which is then analyzed by LLM-based agents. Manipulated log output or external research data could influence the optimization logic. Ingestion points include container logs, experiment history, and knowledge retrieved via Retrieval-Augmented Processing (RAP). The runner performs sanitization of metadata and identifiers in the feedback loop, and the skill validates action specifications against structured JSON schemas. Capability inventory includes shell command execution via various SDKs, network requests to LLM APIs, and workspace file writing.
- [EXTERNAL_DOWNLOADS]: The skill installs Python packages and pulls container images from official NVIDIA repositories. These resources trace back to the author's infrastructure and are used for the standard operation of the AutoML workflow.
- [COMMAND_EXECUTION]: The skill executes shell commands to perform preflight checks, install dependencies, and manage training jobs through platform-specific SDKs such as Docker, Slurm, and Kubernetes.
- [DYNAMIC_EXECUTION]: The skill supports Python hooks and local process execution, allowing for custom metric extraction and evaluation logic during the optimization loop.
Audit Metadata