What to do When Your AI Guardrails Fail

A Microsoft 365 Copilot bug (CW1226324) allowed the AI to read and summarize confidential emails for weeks despite sensitivity labels and DLP policies being correctly configured to block it. The incident exposes a critical architectural flaw: all AI governance controls lived inside the same platform as the AI itself, creating a single point of failure with no independent detection layer. The author argues organizations must stop trusting AI platforms to govern themselves and instead implement defense-in-depth at the data layer — independent governance layers, purpose binding, least-privilege controls, and audit trails the organization controls. Compliance exposure under GDPR Article 32 and the EU AI Act is also discussed.