How NuBench AI protects your data

The life of a prompt

1. 1

   Classification

   Every prompt is classified into one of four tiers before anything else happens. The tier is shown to the user on the message — classification is visible, not a black box.

2. 2

   Policy & DLP

   Policy rules run against the content. Depending on your configuration, matches are flagged (recorded but allowed), redacted (replaced with a marker before leaving your boundary), or blocked outright with an explanation. The dedicated classifier service with expanded PII/PHI detection is in development; current workspaces run the pipeline’s rule-based checks.

3. 3

   Dispatch under your terms

   The request goes to your model provider — Amazon Bedrock inside the AWS boundary by default, or your own OpenAI, Anthropic, or Azure OpenAI relationship with your key. Provider keys live in AWS Secrets Manager, encrypted at rest, and are read only at dispatch time.

4. 4

   Response inspection

   Responses pass back through classification and DLP on the way in, so the pipeline governs both directions.

5. 5

   Audit

   The exchange, its tiers, and every policy decision are written to the audit trail as hash-chained records — each record’s hash covers the previous one, so tampering anywhere in the chain is detectable.

Commitments

• Customer data is not used to train models — by NuBench or by your provider under BYO terms you control.
• Strong tenant isolation: your data is logically separated from other customers. Dedicated single-tenant deployment — exclusive databases, cache, and buckets — is available at the Enterprise tier.
• Retention for conversation logs is configurable per organisation to match your records policy.
• Files upload directly to your workspace’s encrypted storage and are DLP-scanned before any model sees their content. (File pipeline in development.)
• Exports preserve structured audit records for your SIEM, GRC tooling, or a regulator.