Block, Warn, Mask, Audit — and when each mode fits

Four protection modes, four responses. A pragmatic decision aid per data type and domain.

An allow/deny model is the simplest policy framework — and in practice almost always wrong. Why? Because sensitive data has context.

An IBAN in your internal accounting app? No problem. The same IBAN in a public AI chat? Big problem. The same IBAN in an approved AI tool with DLP protection? Acceptable.

Four modes give you the flexibility that reality demands.

Block — when the damage would be irreversible.

Use Block sparingly. Only where the damage cannot be undone.

Examples: authentication tokens, payroll data, client master data in unauthorized tools. The action is stopped. The incident is logged.

Caution: blocking everything trains employees to find paths outside the browser. WhatsApp. Email. USB drive.

Warn — when context matters.

A contextual warning appears. The person can proceed or cancel.

Ideal for edge cases. Customer names that appear in legitimate workflows — but could also be problematic in other contexts. The decision stays with the human.

Mask — when the workflow should continue.

Sensitive values are replaced with tokens before transmission.

The prompt works. The model gets the structure — without seeing the contents. IBAN becomes [IBAN], name becomes [NAME].

Mask is the most underrated mode. It resolves most real conflicts between productivity and protection.

Audit — when you want to understand first.

Actions continue normally. They are logged.

Audit is ideal for the first two weeks of any new use case. You collect data before you intervene. You understand patterns before you set rules.

Always start with Audit. Then: Mask for routine cases. Then: Warn for edge cases. Then: Block for critical cases.