scenario-orbit-views
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute several local shell commands to facilitate the 3D turnaround process. These include running Blender in background mode with script arguments (
render_grounded.py) to generate layouts and invoking Python scripts for silhouette matching (match_view.py), layout composition (compose_layout.py), and shadow-aware background removal (studio_shadow.py). - [EXTERNAL_DOWNLOADS]: The agent is directed to download 3D assets (GLB files) from the Scenario platform using
curl -Lfollowing the completion of image-to-3D model runs. This is a standard part of the workflow for managing assets across the MCP server and the local environment. - [INDIRECT_PROMPT_INJECTION]: The skill represents an indirect injection surface because it ingests untrusted user content, such as a source image and various descriptive parameters (identity, materials, colors, and ground contact). This data is interpolated into a multi-paragraph "Repaint" prompt. While the instructions provide strong contextual framing, there are no explicit boundary markers (like XML tags or delimiters) used to isolate the user-provided text from the system instructions.
Audit Metadata