sinch-voice-api-v2
Pass
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: Defensive instructions against Indirect Prompt Injection. The skill contains specific guidance to treat all external data—including webhook bodies, caller input, and transcribed speech—as untrusted. It explicitly mentions that phrases like 'ignore previous instructions' appearing in transcripts should be treated as data rather than agent commands.
- [SAFE]: Secure credential management. The skill provides clear instructions and examples for storing sensitive project IDs, key IDs, and secrets in environment variables, with a strong warning against hardcoding credentials in source code or commands.
- [SAFE]: Controlled documentation fetching. The skill follows a 'fetch-first' policy for technical details (Tier B facts) but restricts this activity to trusted first-party vendor domains (developers.sinch.com and dashboard.sinch.com). This ensures the agent uses authoritative, up-to-date schemas without risk of following malicious external links.
- [SAFE]: Operation classification and consent. The skill implements a binding policy requiring the agent to classify operations as read-only, billable, or destructive, and to obtain explicit user approval before performing billable or destructive actions.
Audit Metadata