The Agent Skills Directory

Dynamic Script Execution: The skill enables the execution of Three.js code via <script type="text/three">. This requires the allow-client-scripts attribute, which acts as a security gate. The execution is isolated within a Web Worker or a Rust-based sandbox (Boa), which prevents direct DOM access and revokes network APIs (fetch, XMLHttpRequest) to mitigate data exfiltration risks.
Indirect Prompt Injection Surface: The connectGeminiLive and holdToAsk features ingest user voice data to interact with multimodal models. This introduces a potential surface for indirect prompt injection. However, the skill documentation highlights the use of sandboxing and specific event handling to manage these interactions safely.
External Resource Loading: The skill describes loading 3D models and script bundles from external URLs. It provides clear guidance on configuring Content Security Policy (CSP) headers, such as connect-src and worker-src, to ensure these remote resources are loaded securely and to prevent unauthorized connections.
Spatial State Persistence: The persistent-anchors and persist-url-hash features allow saving 3D scene layouts to localStorage or URL hashes. The framework enforces an 8KB limit on these tokens and includes validation to prevent prototype pollution attacks via the persisted state.
Simulation and Tooling: The inclusion of three-live-cli and various MCP tools allows for headless validation and testing of scenes, which can be used to verify the integrity and safety of authored content before deployment.

three-live-authoring