character-autorig
Pass
Audited by Gen Agent Trust Hub on Nov 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONOBFUSCATIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external 3D mesh files (.glb, .gltf, .obj) and JSON marker files. This creates a surface where instructions could be embedded in data processed by the agent.
- Ingestion points: Mesh data is loaded and processed in
mesh_check.pyandarp_markers.py. - Boundary markers: Absent; the skill relies on the structure of the 3D files.
- Capability inventory: The skill can execute shell commands via Blender and write files to the local directory (JSON reports and PNG renders).
- Sanitization: The skill converts geometry to numeric arrays (numpy), which effectively sanitizes geometric data, but it does not explicitly sanitize textual metadata from mesh nodes.
- [COMMAND_EXECUTION]: The skill uses
subprocess.runandsubprocess.Popenin scripts likerig_exam.pyandautorig.pyto coordinate the execution of Blender and other internal scripts. These calls use argument lists rather than shell strings, which is a standard and safer implementation for orchestration. - [OBFUSCATION]:
mesh_check.pyuses thebase64module to decode data URIs within glTF files. This is a standard operation for 3D model parsing and is not used to hide malicious logic. - [DYNAMIC_EXECUTION]:
arp_smart.pyuses__import__to dynamically load the Blendermathutilsmodule. This is necessary for Blender-specific Python scripts and does not involve executing untrusted code.
Audit Metadata