model-pass
Pass
Audited by Gen Agent Trust Hub on Sep 27, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes local shell scripts to automate 3D rendering audits and image comparisons, which are essential for its stated purpose.
scripts/shot.shinvokes theagent-browsertool to capture screenshots of the 3D studio, setting specific browser flags like--disable-gpu-vsyncand--disable-frame-rate-limitto ensure stable headless rendering.scripts/compare.shemploys themagick(ImageMagick) utility to perform image montages and pixel-level comparisons (magick compare -fuzz 3%) to verify that model changes do not introduce unintended visual regressions.- [INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for processing 3D model data, which acts as an ingestion surface for potentially untrusted external content.
- Ingestion points: The skill processes 3D model files (geometries, materials) and project-specific configuration files like
3D-STYLE.mdto perform its audits. - Boundary markers: While the skill uses specific studio helpers to delimit measurements, it lacks explicit prompt-level delimiters for data extracted from model metadata.
- Capability inventory: The agent is granted capabilities for browser interaction, screenshot capture, and local command execution related to image processing.
- Sanitization: The audit process includes geometric validation via
assets/studio-helpers.ts, which checks for non-finite values (NaN) and zero-length normals to prevent rendering artifacts and bloom-related 'black frame' issues.
Audit Metadata