Data-loss prevention

Every prompt and every response passes through classification and DLP steps in the governance pipeline — before leaving your boundary on the way to a model, and again on the way back.

Enforcement status: classification and DLP scanning are part of the pipeline design and run as rule-based checks. The dedicated classifier service with expanded PII/PHI detection is in development; current workspaces run the pipeline's built-in rule set.

Classification tiers

Content is classified into four tiers:

| Tier | Typical content | | --- | --- | | PUBLIC | Material already public or intended to be | | INTERNAL | Ordinary business content | | CONFIDENTIAL | Client work product, financials, personnel matters | | RESTRICTED | Regulated data, privileged material, secrets |

Tiers are visible to users on every message, so classification is not a black box — your team sees what the platform saw.

Policy actions

For each rule match, policy determines the action:

• Flag — allow the request, record the finding in the audit trail.
• Redact — replace the matched content with a [REDACTED:rule-name] marker before it leaves your boundary. Downstream services and audit logs see the redaction marker, never the original value.
• Block — refuse the request entirely, with a clear explanation to the user.

Configuring DLP

From the admin console, organisation and IT admins set:

1. The DLP mode (off / flag / block) as the workspace default.
2. Rule sensitivity for common categories: identifiers, financial data, credentials.
3. Tier-based routing constraints — for example, RESTRICTED content may be confined to specific providers or blocked from AI entirely.

Every DLP decision — match, action, rule name — is written to the audit trail so compliance can evidence the control, not just assert it.