blender
Pass
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input data, such as project briefs and narrations, to generate structured JSON 'shot specs'. These specs are then interpreted by a local Python tool to control Blender's rendering engine, creating a vulnerability surface where malicious instructions in the source text could attempt to manipulate the resulting 3D scene or file operations.
- Ingestion points: Processes 'briefs' and 'narration' (via captions) to construct the
spec.jsonfile which drives the automation. - Boundary markers: The skill does not document the use of specific delimiters or 'ignore embedded instructions' warnings when the agent interpolates user-provided content into the shot spec.
- Capability inventory: The
tools/blender_render.pyscript has the capability to read local files (3D models, textures, SVG assets) and write data to the filesystem (renders, events, and Blender project files). It also invokes the Blender executable as a subprocess. - Sanitization: There is no evidence of path sanitization or strict schema validation to prevent a malformed spec from targeting sensitive local files outside the project directory.
- [COMMAND_EXECUTION]: The skill relies on executing local Python scripts (
tools/blender_render.py) usinguv run, which in turn launches Blender in a headless mode to execute dynamically generated Python scripts (-P script.py). This is the intended primary purpose of the skill for 3D automation.
Audit Metadata