product-launch-video
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests content from external URLs using the
capturecommand. This content is used to generate the video storyboard and script, creating an attack surface where malicious website data could influence the agent's instructions. Evidence: (1) Ingestion point: Step 1 in SKILL.md useshyperframes capture. (2) Boundary markers: The skill references a 'brief contract' and uses an intent layer for validation. (3) Capability inventory: Subprocess execution (ffmpeg, node), file writing, and sub-agent dispatch. (4) Sanitization: Basic structural validation exists in the assembly scripts, but no semantic filtering of captured text is performed. - [DYNAMIC_EXECUTION]: The skill programmatically generates HTML and JavaScript files (index.html and frame compositions) for animation. Scripts like
transitions.mjsandassemble-index.mjsinject GSAP code and transition logic into these project files. - [COMMAND_EXECUTION]: Several local scripts utilize
spawnSyncto execute system tools such asffmpeg,ffprobe, andnodefor media processing and task automation. These calls use array-based arguments to prevent shell injection. - [EXTERNAL_DOWNLOADS]: The generated output includes the GSAP library loaded via a public CDN (jsdelivr.net). The skill also interacts with vendor APIs (HeyGen) for audio processing and asset retrieval.
Audit Metadata