Agent Sandbox Taxonomy (AST) — Evaluation & Composition Skill

To evaluate AI agent sandbox security, you score products against the Agent Sandbox Taxonomy (AST) framework, assessing 7 defense layers and 7 threat categories to generate a detailed scorecard and threat coverage fingerprint.

The best way to compose AI sandboxes is to stack complementary products based on your threat model, using AST composition rules to fill security gaps across compute isolation, network controls, and credential management for comprehensive threat coverage.

Sandbox scoring works by evaluating defense layers like compute isolation (L1), network controls (L4), and credential management (L5) against the AST framework to prevent data exfiltration and destructive operations when AI agents interact with sensitive cloud resources.

Yes, you can use the AST framework to assess Docker sandbox products. It scores Docker against strength, granularity, and portability dimensions across 7 defense layers to determine its suitability for secure AI agent deployment.

You need to combine multiple sandbox products when a single product leaves security gaps in your threat model. The AST framework identifies these gaps and recommends stacking complementary sandboxes to achieve full defense layer coverage.

The Agent Sandbox Taxonomy evaluates defense layers using three scoring dimensions: strength, granularity, and portability. These dimensions assess how robustly each sandbox product mitigates threats across the 7 defined threat categories.