whitebox-previs-executor
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes local binaries including Blender, FFmpeg, and FFprobe to perform 3D rendering and video encoding. These operations use the
subprocessmodule with argument lists, which avoids shell-based command injection vulnerabilities. The skill also implements isolation by redirecting environment variables likeTEMPandLOCALAPPDATAto a local runtime directory. - Evidence:
scripts/run_blender_previs.py,scripts/run_blender_rigged_fight.py, andscripts/render_previs.pycontain logic for discovering and invoking these external tools. - [INDIRECT_PROMPT_INJECTION]: The skill processes natural language video prompts from the user to generate 3D scene specifications. While this creates a potential surface for indirect injection, the skill employs strict schema validation and does not use the processed content in a way that would trigger further agent actions or privilege escalation.
- Ingestion points: Natural language video descriptions provided by the user to the agent.
- Boundary markers: Skill instructions direct the agent to transform input into the structured
previs-compiler/1.0JSON schema, acting as a structural boundary. - Capability inventory: Ability to write media files (MP4, .blend) and execute rendering software via subprocesses.
- Sanitization: Generated JSON data is comprehensively validated by
scripts/validate_compiled_previs.pyandscripts/validate_spec.pybefore any rendering or execution occurs.
Audit Metadata