taste-application
Pass
Audited by Gen Agent Trust Hub on Sep 11, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the
subprocessmodule to execute external binaries including FFmpeg, ffprobe, and Blender for media processing and 3D rendering tasks. - Evidence: Found in
scripts/taste/assemble.py,scripts/taste/frames.py, andscripts/taste/render3d.pyfor task automation. - Context: These operations are essential to the skill's purpose and include extensive safeguards like input validation, absolute path resolution, and restricted execution environments.
- [EXTERNAL_DOWNLOADS]: Fetches generated video clips and 3D assets from the
fal.mediadomain over HTTPS. - Evidence: Implemented in
scripts/falapi.pyviaurllib.requestusing a customSafeRedirecthandler. - Context: The implementation enforces host validation to ensure downloads only originate from authorized and well-known service providers.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text and media metadata to construct prompts for remote AI models (VLM and video generators).
- Evidence: Prompt construction logic in
scripts/apply.pyandscripts/distill.pyinterpolates user strings. - Context: It mitigates injection risks by using structured JSON schemas for model outputs, explicit boundary markers (WHAT/HOW/GRADE), and secondary validation steps.
- [DYNAMIC_EXECUTION]: Dynamically locates and loads the DaVinci Resolve scripting module and executes scripts within Blender.
- Evidence:
load_resolve_moduleinscripts/resolve_ingest.pyandturntable_blenderinscripts/taste/render3d.py. - Context: This facilitates professional NLE integration and is restricted to documented vendor installation directories and temporary local scripts.
- [DATA_EXFILTRATION]: Performs network operations to interact with model endpoints and upload reference stills.
- Evidence: API interaction layer in
scripts/falapi.pyusingfal-clientandurllib. - Context: Security is managed through environment-based credential handling (
FAL_KEY), explicit redaction of signed URLs and secrets in logs, and a 'fail-closed' policy requiring an explicitTASTE_FORGE_ALLOW_LIVE=1flag.
Audit Metadata