talking-avatar-skill
Pass
Audited by Gen Agent Trust Hub on Sep 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill implements a secure 'Bring Your Own Key' (BYOK) flow where OpenAI API keys are handled only in memory and never persisted in local storage or browser cookies.
- [SAFE]: All network communication is restricted to official OpenAI domains (api.openai.com) and the application's own backend routes for session negotiation.
- [SAFE]: Scaffolding and asset validation are performed by bundled local Python scripts that only execute authorized file transformations and dimension checks within the project directory.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input (speech and text). 1. Ingestion points: User speech and text input via TalkingAvatarApp.tsx. 2. Boundary markers: Explicit system instructions in route.ts.tmpl forbid the model from revealing sensitive information. 3. Capability inventory: Limited to conversation via OpenAI's Realtime API with no access to local system resources based on user input. 4. Sanitization: Relies on upstream provider (OpenAI) safety guardrails.
Audit Metadata