Trust Center
The architecture, controls, and posture behind the platform — written for the people who run your security review.
No. NuBench does not train models on customer data. Under the default Amazon Bedrock integration, AWS does not use customer content to train models. Under bring-your-own-key, your data-processing terms with your provider apply — which you control.
No. Each customer workspace is a single-tenant AWS deployment with its own database, cache, storage, and encryption keys. Your data shares no data store with another customer.
In transit, TLS on every public endpoint and required TLS to the database. At rest, every data store is encrypted with a dedicated customer-managed KMS key with annual rotation — six separate keys covering the database, cache, object storage, logs, secrets, and infrastructure state.
Access follows least privilege: each service has an IAM role scoped to exactly what it needs, and human access to production is governed by the same IAM discipline. Every decrypt operation on a customer-managed key is recorded by AWS CloudTrail, so access to encrypted data leaves durable evidence.
In AWS, in your workspace’s deployment region (US East today; additional regions are a procurement conversation). Data does not move between regions as part of normal operation.
The prompt is classified and scanned before it leaves your boundary. Depending on your policy, sensitive matter is flagged and recorded, redacted with a visible marker, or the request is blocked with an explanation. Every decision is written to the audit trail with the rule that fired.
Yes — OpenAI, Anthropic, and Azure OpenAI keys are supported alongside the default Amazon Bedrock integration. Keys are stored in AWS Secrets Manager under a customer-managed KMS key and read only at dispatch time.
Budgets are defined in tokens at organisation, matter (cost code), project, and optional user scope, and evaluated in the request pipeline before any model is called. Hard caps refuse overspend before dispatch; Warning and Critical thresholds remain advisory in the consoles. If the budget counter store is unavailable, requests fail closed rather than running uncapped.
Yes. Every prompt, response, policy decision, and administrative action is written as a structured, hash-chained record — alteration anywhere in the chain is detectable. Auditors get read-only access, and exports preserve the structured format for your SIEM or GRC tooling.
Not yet, and this page will not pretend otherwise. The platform is built on SOC 2-relevant controls (immutable audit logging, least privilege, reviewed change management, encryption everywhere) and a formal attestation is on the roadmap. Ask us for the current control mapping.
AWS provides the infrastructure. Your model provider is a subprocessor only in the configuration you choose: Amazon Bedrock keeps requests inside AWS; bring-your-own-key routes to the provider you already contract with, under your terms.
All infrastructure and application changes flow through version control, code review, and CI with static security analysis. Production deployments are automated with automatic rollback on failure. There are no manual production changes.