agent-ui

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs the user to download component definitions from https://ui.inference.sh/r/agent.json using the shadcn CLI. It also requires the installation of the @inferencesh/sdk package from the NPM registry. These resources are official components provided by the vendor to enable the skill's agentic UI functionality.
  • [INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of agents that can scan and interact with application UI elements (e.g., using scan_ui and fill_field tools). This capability introduces a surface where instructions embedded within the application's user interface could potentially influence agent behavior.
  • Ingestion points: Client-side UI scanning and form-filling tools defined in the Agent component.
  • Boundary markers: The skill relies on the underlying LLM's system prompt and the developer's implementation of the SDK to maintain execution boundaries.
  • Capability inventory: The agent can interact with browser UI elements, fill forms, and execute client-side tools.
  • Sanitization: Sanitization and validation of UI-extracted data are dependent on the implementation of the provided SDK and the developer's integration logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 05:54 PM
Security Audit — agent-trust-hub — agent-ui