AI is handing CISOs a new question to own: which data may each model and agent ever see. The CISOs who answer it with classification will be the ones enabling AI rather than blocking it.

Germany's intelligence services reportedly plan an air gapped cloud on Google technology. Isolation answers where data sits. Sovereignty also asks who controls the software, the updates and the keys, and classification decides which data needs which answer.

525,000 people in one university breach shows how much personal data research institutions hold, often unclassified and spread across systems. Knowing which records are sensitive, and protecting those first, keeps an incident from becoming a crisis.

Air gapped and hosted in Germany answers the location question. The harder one is control: who decides which data reaches the highest classification, and who holds the keys when software and updates come from abroad. Sovereignty is fundamentally decided at the data layer.

This shift makes data sovereignty the primary pillar of defense.

Precisely, most email systems have AI ingestion.

PRODUCT RELEASE: Réunion version 3.3.3 of the Android app is now available on the Play Store for upgrade/download. changelog: - Admin recovery: restore a profile on a new phone with the admin's help. - New approval engine: recovery and escrow requests in one Approvals inbox. - Escrow officers can approve or decline requests from their android phone (previously iOS exclusive) - Redesigned EasyCrypt interface: new folders, mail list, message and composer screens and tons of other visual items. - EasyCrypt signatures, folder management and editable Outbox mail. - Lots of small fixes and reliability improvements. Thanks to everyone that have reported issues. As usual, when you run into issues please remember to provide me with your Installation ID (Réunion App > Settings > App Usage Statistics) and details about the problem so we can check the logs and understand what happened.

Preparing for disruption starts with three questions: 1. If your cloud goes dark, who still holds your mail? 2. Does sensitive data stay under your own jurisdiction? 3. Does classification survive the fallback channel? Resilience and sovereignty are one conversation.

European, but also Japanese Governments seeking sovereign AI still face the same leak path: unclassified sensitive data in email and attachments that models can read before any national control is applied.

Limiting agent permissions is necessary, yet the earlier control is classifying the email and its attachments so the agent never sees material it is not allowed to act on.

Residency rules apply equally to the content that ends up in logs; classifying sensitive payloads before they generate those logs keeps the entire chain under sovereign control.

Detection speed is useful; for regulated communications the control is classifying attachments and content before any system, sovereign or not, can ingest them.

I think the aim is to achieve both the world-class AI orchestration technology they have been advocating and a domestic LLM that can process sensitive information without depending on overseas models. Aiming for both “performance” and “no data leakage.”

The tension is resolved by classifying content before it reaches any model, so usage stays productive while sensitive data never leaves the sovereign boundary.

Classification and sovereign control over what models can access matter more than after an incident sandboxing when agentic systems can act outside their stated reasoning.