explainer-video

Warn

Audited by Gen Agent Trust Hub on Oct 26, 2026

Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources including web pages, PDF documents, Word files, and git repositories to generate narration and video content.
  • Ingestion points: scripts/source/snapshot.mjs reads content from user-provided URLs and file paths into local snapshots.
  • Boundary markers: The skill dispatches sub-agents (agents/source-reader.md, agents/researcher.md) with explicit instructions to "change no word" and "ignore previous instructions" (as a principle of verbatim copying), though static analysis cannot guarantee the effectiveness of these instructions against adversarial input.
  • Capability inventory: The skill possesses extensive system-level capabilities, including executing shell commands (ffmpeg, git, uv, npm), writing to the filesystem, and making authenticated network requests to AI providers.
  • Sanitization: The snapshotting process includes specific logic to withhold files or text segments matching common secret patterns (e.g., API keys, private keys).
  • [EXTERNAL_DOWNLOADS]: The scripts/setup.mjs utility automates the retrieval of binary tools, machine learning models (vWav2Vec2, Kokoro, ACE-Step), and font files from numerous external repositories and CDNs including GitHub, HuggingFace, JSDelivr, and Microsoft's NuGet gallery.
  • [DYNAMIC_EXECUTION]: The skill manages a per-user tool cache and executes scripts from dynamically computed paths. It uses node -e to probe for CUDA availability and runs Python scripts (scripts/voice/vc/vc.py, scripts/score/ace.py) through uv run using local virtual environments.
  • [COMMAND_EXECUTION]: The skill frequently invokes system utilities such as ffmpeg and ffprobe for media manipulation, git for repository analysis, and uv or npm for dependency management via spawn and spawnSync calls.
  • [CREDENTIALS_UNSAFE]: Operation requires the user to provide sensitive environment variables (GEMINI_API_KEY, ELEVENLABS_API_KEY) for third-party AI synthesis services.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 26, 2026, 08:48 AM