tao-generate-referring-expressions

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests and processes untrusted external data (images and text-based label files) which are then interpreted by a Vision-Language Model (VLM).
  • Ingestion points: The skill reads images from data.image_dir and KITTI-format label files from data.kitti_label_dir (referenced in SKILL.md and references/configuration.md).
  • Boundary markers: The documentation does not specify the use of delimiters or 'ignore' instructions to prevent the VLM from following commands embedded within images (OCR-based injection) or label text.
  • Capability inventory: The skill executes the auto_label CLI tool within a Docker container and performs network operations to external VLM endpoints (Gemini or OpenAI-compatible APIs).
  • Sanitization: There is no evidence of sanitization or validation of the contents of the image or label files prior to processing by the VLM.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:56 AM
Security Audit — agent-trust-hub — tao-generate-referring-expressions