claude-video-enhance-audio

Pass

Audited by Gen Agent Trust Hub on May 15, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses shell variables like $INPUT and $OUTPUT directly within Bash tool blocks (e.g., python3 scripts/audio_enhance.py separate "$INPUT"). While the use of double quotes provides some protection, this pattern remains a potential surface for command injection if input filenames contain sophisticated shell metacharacters or if the agent environment performs unsafe expansion.
  • [PROMPT_INJECTION]: This skill is vulnerable to Indirect Prompt Injection (Category 8). It processes untrusted audio/video files to generate transcripts and speaker labels (via WhisperX and pyannote-audio).
  • Ingestion points: External media files provided as $INPUT in SKILL.md.
  • Boundary markers: None identified. The transcript output is processed by the agent without explicit instructions to ignore embedded commands.
  • Capability inventory: The skill has access to the Bash, Read, and Write tools, allowing for command execution and file system modification.
  • Sanitization: No evidence of sanitization or filtering of the generated transcript text before it is returned to the agent context.
  • [CREDENTIALS_UNSAFE]: The skill instructions suggest passing sensitive API tokens (HuggingFace, ElevenLabs, OpenAI) via command-line flags (e.g., --hf-token TOKEN). This is a sub-optimal security practice as it may expose credentials in process lists or shell history logs, although the skill also correctly mentions the use of environment variables.
Audit Metadata
Risk Level
SAFE
Analyzed
May 15, 2026, 02:55 PM
Security Audit — agent-trust-hub — claude-video-enhance-audio