narova-xiaomi
Pass
Audited by Gen Agent Trust Hub on Sep 1, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill transmits synthesis text and, optionally, local audio reference files to Xiaomi MiMo API endpoints (api.xiaomimimo.com). The transmission is the primary intended function of the TTS provider and is conducted over encrypted HTTPS connections. The worker implements safety limits on input text length (8,000 characters) and audio reference file size (10MB) to mitigate potential resource abuse.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text and delivery instructions which are interpolated into a request for a third-party AI model. Ingestion points: The worker accepts text, instructions, and voice design parameters via JSONL on stdin in tool/worker.py. Boundary markers: It utilizes a structured message format with defined user and assistant roles to separate control instructions from the content to be synthesized. Capability inventory: Includes the ability to read local audio files (for voice cloning), write WAV files to disk, and perform network POST requests to the synthesis API. Sanitization: The skill validates that file paths are absolute regular files (not symlinks) and enforces character limits on all inputs to prevent injection or buffer-related issues.
- [SAFE]: The skill implements credential redaction logic to ensure that the required MIMO_API_KEY environment variable is never exposed in error messages or stdout/stderr logs. Setup procedures involve only local script execution and standard library checks, avoiding unverifiable third-party dependencies.
Audit Metadata