AI Security Must Go Beyond the Model: CertiK Identifies Google EdgeTPU Vulnerabilities, Highlighting New Risks in AI Infrastructure

## Executive Summary CertiK disclosed two EdgeTPU vulnerabilities (CVE-2026-0150 — High; CVE-2026-0153 — Critical) that may let attackers bypass Android–EdgeTPU isolation, execute unauthorized operations on the privileged inference chip, or access on-device data; the report uses this case to argue that AI security must move beyond model-focused protections to secure the full system and lifecycle.