build-avatar-app
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The documentation in
references/server.mdprovides a command to install theatmee-toolkitCLI by downloading and executing a shell script from the author's official GitHub repository (https://raw.githubusercontent.com/Atmanity/atmee-toolkit/main/install.sh). This is an intended setup procedure for the vendor's developer tools. - [EXTERNAL_DOWNLOADS]: The
templates/try.htmlfile references thelivekit-clientlibrary via the well-known JSDelivr CDN (https://cdn.jsdelivr.net/npm/livekit-client@2.22.3/+esm). This is a standard practice for including reputable third-party client-side dependencies. - [INDIRECT_PROMPT_INJECTION]: The skill involves building an interface that processes real-time user input from a LiveKit room, which represents a potential attack surface for instructions embedded in user speech or text.
- Ingestion points: Untrusted user data enters the system through the LiveKit room's transcription (
lk.transcription) and chat (lk.chat) topics, as handled intemplates/avatar-room.js. - Boundary markers: The architecture uses structured
stateandeventmessages to communicate context to the avatar. While the templates demonstrate sending data, they do not explicitly define developer-level delimiters or safety instructions to separate system context from raw user-provided strings. - Capability inventory: The AI avatar has the capability to trigger "app tools" (client-side functions) such as updating the user interface (
show_slide) or fetching dynamic data (get_score). - Sanitization: The provided example code does not include explicit sanitization or filtering of transcription data before it is processed by the AI agent's runtime environment.
Audit Metadata