building-html-canvases
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The documentation describes best practices for developing within a sandboxed React environment. It correctly identifies and reinforces security boundaries, such as the network origin allowlist (capabilities.network.origins) and the requirement for explicit user interaction for external links via a restricted bridge (ph.openExternal).
- [INDIRECT_PROMPT_INJECTION]: The skill defines a framework where canvases ingest data through the ph.query bridge, establishing an indirect injection surface. The instructions mitigate this by prescribing transparency measures, such as disclosing raw queries in the UI.
- Ingestion points: Data retrieved from the ph.query bridge (referenced in SKILL.md).
- Boundary markers: The documentation requires a Verifiability pattern, where data-backed claims include links to insights or disclosure of the underlying query in a element.
- Capability inventory: ph.query (data access), ph.capture (event tracking), ph.openExternal (navigation), and low-level browser APIs (Canvas2D, WebGL, Web Audio).
- Sanitization: Refers to platform-level validation tools described in building-canvases and related skills.
Audit Metadata