hung-yi-lee

Pass

Audited by Gen Agent Trust Hub on Aug 12, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill provides Python scripts (scripts/hungyi_kb.py) and instructs the agent to invoke them via the shell to manage and query the knowledge base (e.g., python3 scripts/hungyi_kb.py search "<query>"). This pattern allows the agent to execute shell commands where input strings might not be strictly sanitized, leading to potential command injection.
  • [EXTERNAL_DOWNLOADS]: The management scripts use well-known tools like yt-dlp and the youtube-transcript-api to fetch metadata and transcript files from YouTube. While YouTube is a well-known and reputable service, this constitutes an external network dependency during maintenance tasks.
  • [PROMPT_INJECTION]: Educational transcripts included in the knowledge base (such as raw/youtube/transcripts/dWQVY_h0YXU.md) contain explicit examples of prompt injection and jailbreak techniques. This creates a surface for indirect prompt injection:
  • Ingestion points: Untrusted transcript data is read from the raw/youtube/transcripts/ directory during search and query operations.
  • Boundary markers: The content is processed as standard Markdown and lacks robust boundary markers to instruct the agent to ignore instructions embedded within the transcripts.
  • Capability inventory: The skill possesses shell execution capabilities via its provided Python utilities.
  • Sanitization: No explicit sanitization or filtering of the transcript content for malicious instructions was observed.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 12, 2026, 11:00 AM
Security Audit — agent-trust-hub — hung-yi-lee