twinmind-performance-tuning

Pass

Audited by Gen Agent Trust Hub on Apr 4, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill performs network requests to the legitimate TwinMind API (api.twinmind.com) for health checks and transcription. It also includes functionality in references/implementation.md to fetch audio data from arbitrary URLs for the purpose of content hashing and transcription.
  • [COMMAND_EXECUTION]: The skill utilizes the fluent-ffmpeg Node.js library in references/implementation.md to process audio files. This involves executing the ffmpeg system binary to perform operations like noise reduction, normalization, and format conversion.
  • [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface through its audio transcription workflow.
  • Ingestion points: Untrusted audio data enters the agent context via URLs processed in the transcribeWithCache function in references/implementation.md.
  • Boundary markers: There are no explicit delimiters or protective instructions provided to prevent the agent from following commands that might be present in the resulting transcript.
  • Capability inventory: The skill utilizes network access (fetch, curl), file operations (Read, Write, Edit), and shell execution via FFmpeg.
  • Sanitization: The implementation does not show any sanitization or validation of the transcribed text before it is returned to the agent context.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 4, 2026, 07:31 PM