tao-analyze-gaps-od-map

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFECOMMAND_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute TAO Data Services gap analysis by running a Docker container (nvcr.io/nvidia/tao/tao-toolkit). This involves mounting host directories (RUN_ROOT) and using host networking/IPC settings to perform heavy-duty data processing.
  • [PRIVILEGE_ESCALATION]: The documentation provides a command to fix file ownership issues by running an Alpine container to chown the output directory. While this is a standard mitigation for root-running containers, it involves using privileged Docker access to modify host file system permissions.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external annotation files (KITTI or COCO formats) provided in ground_truth_ann_path and inference_ann_path. These files represent a surface for indirect prompt injection if the processing engine inside the container is vulnerable to malformed data or embedded instructions.
  • Ingestion points: Annotation files specified in the YAML configuration (ground_truth_ann_path, inference_ann_path).
  • Boundary markers: None identified in the skill instructions or validation script.
  • Capability inventory: The analysis runs inside a Docker container with access to specified host volume mounts.
  • Sanitization: The verify_object_detection_spec.py script validates the configuration structure but does not sanitize the contents of the annotation data files.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:56 AM
Security Audit — agent-trust-hub — tao-analyze-gaps-od-map