baoyu-youtube-transcript
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill spawns external processes using
child_process.spawnSyncto invoke tools such asyt-dlp,bun,uvx, orpython3 -m yt_dlp. These commands are used to fetch video metadata and transcript data. While the input video IDs are validated via regular expressions inscripts/shared.tsto prevent direct command injection, the reliance on external binaries increases the attack surface. - [EXTERNAL_DOWNLOADS]: In
scripts/youtube.ts, the skill invokesyt-dlpwith the--remote-components ejs:githubflag. This flag instructs theyt-dlputility to fetch and load additional extractors or plugins from GitHub, representing a dynamic dependency on external code hosted on a third-party platform. - [INDIRECT_PROMPT_INJECTION]: The skill ingests raw transcript text, video titles, and descriptions from YouTube, which are untrusted external sources. This data is then formatted and passed to an AI model for post-processing (speaker identification and summarization) using the template in
prompts/speaker-transcript.md. The prompt lacks explicit boundary markers or instructions to ignore commands embedded within the transcript, making it susceptible to malicious instructions placed in YouTube captions. - Ingestion points: Raw transcript snippets, video descriptions, and titles fetched via YouTube's InnerTube API or
yt-dlp(scripts/youtube.ts). - Boundary markers: Absent. The prompt template does not provide clear delimiters or warnings to the sub-agent to treat the transcript content as data only.
- Capability inventory: The skill has the ability to write files to the local filesystem (
scripts/storage.ts) and perform network requests to YouTube domains. - Sanitization: The skill performs basic HTML unescaping and tag stripping (
scripts/transcript.ts) but does not sanitize content against natural language instructions.
Audit Metadata