physical-ai-event-video-generation

Pass

Audited by Gen Agent Trust Hub on Sep 2, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes kubectl commands to inspect pod status, query node GPU capacity, and verify environment variables within the scheduler pod. It also uses curl to trigger and monitor Airflow DAG runs via a local ClusterIP. These commands are essential for its primary function of Kubernetes orchestration and workflow management.\n- [EXTERNAL_DOWNLOADS]: The skill interacts with external storage (S3) for uploading seed images and downloading generated datasets. It also connects to user-defined HTTPS endpoints for VLM, LLM, and Image2Video inference services. These downloads are directed to and from locations explicitly provided by the user.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted image data from external URLs or S3 buckets.\n
  • Ingestion points: The input_path parameter in SKILL.md allows the agent to read images from S3 or HTTP/HTTPS sources.\n
  • Boundary markers: While explicit prompt boundaries for image content are not applicable, the workflow isolates image synthesis within the Cosmos3 and labeling containers.\n
  • Capability inventory: The skill has the capability to execute shell commands (kubectl, curl) and perform network operations (aws s3 sync).\n
  • Sanitization: scripts/upload_images.py performs signature and extension validation for JPEG and PNG files, and scripts/payload.py enforces strict Pydantic validation on the workflow configuration.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 2, 2026, 07:16 PM
Security Audit — agent-trust-hub — physical-ai-event-video-generation