tao-analyze-gaps-vlm-bcq

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests model-generated text from a predictions_json file. This represents a potential surface for indirect prompt injection if the VLM output contains malicious instructions that are subsequently processed by an agent.
  • Ingestion points: The predictions_json file processed by the gap analysis tool.
  • Boundary markers: The skill uses structured JSON for data ingestion but does not employ specific delimiters or warnings to ignore instructions within the VLM response fields.
  • Capability inventory: The skill leverages Read and Bash tools to execute containerized analysis and read report files.
  • Sanitization: The analysis logic utilizes word-boundary matching for 'yes' and 'no' keywords, which restricts the potential for command execution or logic subversion during the analysis phase itself.
  • [EXTERNAL_DOWNLOADS]: The skill references an official container image nvcr.io/nvidia/tao/tao-toolkit:7.2.0-data-services. This resource belongs to the skill's authoring organization (NVIDIA) and is consistent with the stated functionality of the TAO Toolkit.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:56 AM
Security Audit — agent-trust-hub — tao-analyze-gaps-vlm-bcq