AI Sandbox Trust Boundary Gap — Microsoft Copilot
MSRC VULN-210492 · 2026-09A Microsoft Copilot bug bounty report closed as "not a vulnerability" that revealed a real gap: the difference between what an AI model refuses to do and what its execution sandbox is actually capable of doing anyway.
Model-level refusals are a policy layer, not a security boundary — tool execution, sandbox capability, and egress control are separate layers that need independent hardening. Reported to MSRC; disclosed publicly after their review window closed with no objection.