mcp-visual-output

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill defines a system where an AI agent generates JSON specifications to be rendered as UI components. While this is an ingestion point for untrusted data, the skill provides comprehensive mitigation strategies:
  • Ingestion points: Data enters via tool responses in the form of structured JSON specs, as seen in the catalog definitions in references/mcp-integration.md.
  • Boundary markers: The skill uses defineCatalog with Zod schema validation to ensure the AI can only produce expected component structures and prop types, preventing the injection of malicious payloads (detailed in rules/mcp-app-setup.md).
  • Capability inventory: Rendering occurs within a sandboxed iframe with restricted permissions, and components are limited to a pre-defined registry, which prevents arbitrary code execution.
  • Sanitization: Explicit rules for Content Security Policy (CSP) are provided in rules/sandbox-csp.md to block unauthorized network requests and script execution.
  • [EXTERNAL_DOWNLOADS]: The skill references standard, public NPM packages for the MCP and json-render ecosystems (e.g., @json-render/mcp, @modelcontextprotocol/sdk). These are legitimate dependencies for the functionality provided.
  • [SAFE]: The skill serves as a security-oriented guide, providing specific warnings against insecure practices such as using wildcard CSP domains or permissive script-src settings, aligning with best practices for AI-driven user interfaces.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:15 AM
Security Audit — agent-trust-hub — mcp-visual-output