Trust and security model
A2UI sends declarative data instead of agent-generated executable code. This narrows the attack surface, but it does not make an agent, message, catalog, component property, URL, or action trusted.
Security boundaries
| Boundary | Control | Remaining risk |
|---|---|---|
| Protocol | Versioned schema validation | Resource exhaustion, parser defects, replay, partial delivery |
| Catalog | Allowlisted component and function definitions | A permitted component may expose unsafe capabilities |
| Renderer | Host-owned native implementations | URL, Markdown, navigation, WebView, and dependency vulnerabilities |
| Action | Server-side authentication and authorization | Stale state, confused deputy, duplicate execution |
| User | Confirmation and clear provenance | Phishing, misleading content, coercive design |
What declarative rendering improves
An agent cannot directly ship arbitrary JavaScript through the A2UI schema. The host chooses which catalog entries exist and how they render. Invalid component names and properties can be rejected before display. These properties make the boundary more auditable than rendering untrusted HTML or JSX.
They are not absolute guarantees. The official project has fixed URL-opening XSS/tab-nabbing and DataModel prototype pollution, demonstrating that renderer code and dependencies remain part of the trusted computing base.
Required controls
- Validate the exact negotiated protocol version.
- Restrict catalogs, component counts, depth, update rate, URLs, and data-model paths.
- Sanitize Markdown and disable raw HTML by default.
- Treat navigation, iframe, clipboard, download, authentication, payment, and file components as elevated capabilities.
- Authorize every consequential action against current server state.
- Keep credentials, tokens, private prompts, and unnecessary personal data out of UI payloads.
- Preserve accessible text fallback and test failure paths.
Use the production security checklist for release review and the security deep dive for the complete threat model.