talking-head-recut
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data from transcript.json to design graphic cards and determine their content. This presents an indirect prompt injection surface where malicious text in the transcript could influence the agent's design choices. Ingestion Point: transcript.json; Capabilities: file system writes and video rendering via hyperframes CLI.
- [COMMAND_EXECUTION]: The skill performs multiple command-line operations using ffprobe for metadata, ffmpeg for audio extraction and encoding, and the npx hyperframes CLI for transcription and rendering.
- [EXTERNAL_DOWNLOADS]: The skill uses npx to execute and update the hyperframes package and its own skill files (npx hyperframes skills update), which involves retrieving and running remote code from established registries.
- [DYNAMIC_EXECUTION]: The agent dynamically generates HTML/CSS fragments for each video card and assembles a master composition at runtime, which is subsequently rendered into an MP4 file using a headless browser.
Audit Metadata