nvidia-dgx-research

Pass

Audited by Gen Agent Trust Hub on Sep 3, 2026

Risk Level: SAFEPERSISTENCEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [PERSISTENCE]: The skill documentation and metadata (_source.yaml) describe a scheduled maintenance process using a macOS launchd agent (~/Library/LaunchAgents/com.flight505.nvidia-dgx-research-refresh.plist). This mechanism is used to trigger a weekly refresh of the documentation catalog via the provided maintenance script.
  • [INDIRECT_PROMPT_INJECTION]: The skill operates by fetching and processing content from external, untrusted sources such as the NVIDIA developer forums and blog posts using WebSearch and WebFetch. This creates an attack surface where malicious instructions embedded in third-party content could potentially influence the agent's behavior.
  • Ingestion points: Community Q&A threads and engineering blog posts retrieved from external websites.
  • Boundary markers: The instructions do not define specific delimiters or instructions to ignore potential adversarial content within the fetched data.
  • Capability inventory: The skill can execute local scripts, use shell commands (grep), and perform network operations.
  • Sanitization: There is no documented validation or sanitization for the content ingested from the forums or blogs.
  • [EXTERNAL_DOWNLOADS]: The maintenance script scripts/update-catalog.sh uses curl to fetch the documentation catalog (llms.txt) from NVIDIA's official documentation domain (docs.nvidia.com). This is a documented process for synchronizing the skill's local reference snapshot with the official source.
  • [COMMAND_EXECUTION]: The skill instructions direct the agent to use shell commands such as grep to search through local reference files and to execute the provided update-catalog.sh bash script for maintenance tasks.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 3, 2026, 12:38 PM