tao-setup

Pass

Audited by Gen Agent Trust Hub on Sep 21, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The scripts/install-codex-agents.sh script downloads the TAO skill bank from the official vendor repository. * Source URL: https://github.com/NVIDIA-TAO/tao-skill-bank.git * Execution method: The codex plugin marketplace add and codex plugin add commands are used to register and install the skill bank as a platform plugin.
  • [COMMAND_EXECUTION]: The skill uses local shell commands to perform host preflight and authentication tasks. * Environment checks: Executes docker info and nvidia-smi to verify GPU support. * Secure Authentication: Registry authentication with docker login is performed by piping the credential from stdin, avoiding exposure in the process table.
  • [INDIRECT_PROMPT_INJECTION]: The skill functions as an orchestrator that reads configuration and metadata from other skills in the repository. * Ingestion points: Reads references/skill_info.yaml from other skills to determine execution parameters like container images and commands. * Boundary markers: The setup process does not define specific prompt delimiters for this metadata ingestion. * Capability inventory: The orchestrated skills have access to Docker execution, Job submission (Slurm/Kubernetes), and S3 storage operations. * Sanitization: No explicit sanitization of ingested metadata is described in the setup files.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 21, 2026, 02:48 AM
Security Audit — agent-trust-hub — tao-setup