vss-build-vision-ai

Pass

Audited by Gen Agent Trust Hub on Oct 24, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes several internal Python and Bash scripts to orchestrate the deployment process, including Docker Compose management and service graph resolution.
  • scripts/check_credentials.sh: Probes official NVIDIA and Hugging Face endpoints to validate API keys before deployment.
  • scripts/normalize_resolved_yml.py and scripts/validate_resolved_yml.py: Perform structural analysis and security sanitization on generated Compose files.
  • scripts/sync_skills.py and scripts/stage_vss_src.py: Manage internal skill registration and environment staging using Git and standard CLI tools.
  • Lifecycle management involves executing docker compose, pkill, and ngc commands to control the containerized environment.
  • [EXTERNAL_DOWNLOADS]: Fetches necessary components from trusted and well-known sources.
  • Downloads the NVIDIA NGC CLI from api.ngc.nvidia.com.
  • References the official uv installer from astral.sh for environment management.
  • Clones secondary blueprint repositories and configuration objects from official NVIDIA GitHub organizations.
  • [CREDENTIALS_UNSAFE]: The skill handles sensitive API keys (NGC, NVIDIA, Hugging Face) required for AI model access.
  • Security is prioritized by implementing a dedicated validation gate that avoids hardcoding secrets and provides clear instructions for safe management via environment variables.
  • Documentation explicitly warns against hardcoding credentials in transcripts or configuration files.
  • [INDIRECT_PROMPT_INJECTION]: The skill exposes an ingestion surface for potentially untrusted data through alert payloads and external configuration.
  • It mitigates the risk of indirect injection by enforcing a trusted environment flag (VSS_AUTO_DEPLOY=true) required to authorize autonomous deployment actions when the input source is untrusted.
  • [PRIVILEGE_ESCALATION]: The deployment process requires specific system configurations (e.g., firewall rules, Docker version pinning) that involve sudo access.
  • The skill implements safe handling by probing for passwordless sudo availability and providing copy-pasteable handoff blocks for the user when manual authorization is necessary.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 24, 2026, 04:21 PM