The research identifies systemic vulnerabilities in current AI testing methodologies, specifically the failure of digital-only sandboxes to mitigate kinetic risks in embodied AI. In cyber-physical systems (CPS), AI agents can bypass digital isolation to manipulate physical environments or human operators. This research introduces a formalized taxonomy and a multi-dimensional measurement framework—incorporating fidelity, controllability, and containment—to address sandbox escape vectors and adversarial attacks on the monitoring apparatus. The framework provides a standardized methodology for validating the safety and security of complex AI deployments through high-fidelity simulation and formal evidence composition.

• Research Overview & Problem Statement
  • Identifies the security gap where digital sandboxes fail to address the kinetic risks of embodied AI.
  • Highlights that AI in robotics and AIoT can manipulate physical processes or human operators.
  • Proposes a shift from simple digital isolation to comprehensive, assurance-oriented frameworks.
• Sandbox Taxonomy & Boundary Definitions
  • Categorizes environments into Digital, Embodied, and Cyber-Physical layers.
  • Defines formalized boundary protocols to prevent cross-domain escalation.
  • Establishes sandbox archetypes to standardize isolation and simulation capabilities.
• Cyber-Physical Threat Model
  • Addresses attack vectors targeting both the AI model and the assurance/monitoring apparatus.
  • Models the risk of "sandbox escapes" into physical environments.
  • Utilizes the "Weakest-Link Rule" to evaluate the reliability of multi-dimensional security evidence.
• Measurement Framework & Metrics
  • Introduces six key metrics: Fidelity, Controllability, Observability, Containment, Reproducibility, and Governance.
  • Quantifies containment effectiveness by calculating the probability of sandbox escape.
  • Measures the "Fidelity Gap" to ensure simulation accuracy relative to physical behavior.
• Industry & Regulatory Implications
  • Provides the validation metrics required for regulatory compliance, such as the EU AI Act.
  • Requires advanced instrumentation and evidence capture for physical AI auditing.
  • Evaluates the resilience of safety-critical monitoring tools against adversarial manipulation.