Share once, reference forever
A developer pastes an API key once during setup. NuBench vaults it automatically. Every subsequent prompt can reference it by name — without the team ever seeing the raw value again.
How it works — Secrets Vault
Credentials captured at the moment they are shared. Never exposed to the model in plaintext again.
When a user pastes an API key, password, or token into the NuBench chat interface, NuBench intercepts it before it reaches the model, stores it in an encrypted vault, and replaces it with a named reference. The credential is gone from the conversation — but still usable, by name, in any future prompt.
The credential exposure problem
In practice, teams using AI tools for development and operations work share credentials constantly. A developer pastes a database connection string to debug a query. An operations engineer shares an API key to troubleshoot an integration. A consultant pastes a client portal password to demonstrate an automation.
In every standard AI platform, those credentials are now part of the conversation history — visible in logs, potentially surfaced in future context, and sent to a model provider under that provider's data-processing terms. Rotation is the correct response, but it almost never happens. The credential lives in the history indefinitely.
NuBench captures the credential before the model sees it and removes it from the conversation record entirely — so the rotation conversation becomes moot.
How it works
- 01
Detect
The NuBench gateway inspects every outbound prompt for credential patterns — API keys, tokens, connection strings, passwords, and other high-entropy secret formats — before the message is dispatched to any model. Detection runs inline, adding no visible latency to the user.
- 02
Encrypt and vault
Detected secrets are extracted from the prompt and stored in a per-user AES-256-GCM encrypted vault. The vault is scoped to the user and the tenant — no other user or tenant can access it. The raw value is never written to conversation history, audit logs, or the model context window.
- 03
Substitute
The extracted secret is replaced in the prompt with a named reference token — for example, <vault:openai-key>. Future prompts can reference stored secrets by name. NuBench resolves the reference at dispatch time and injects the value directly into the tool call or model context, without surfacing it to the user session or the conversation record.
Technical specification
- Encryption
- AES-256-GCM with a per-tenant KMS-managed key (AWS KMS). Envelope encryption: a data key is generated per secret, encrypted by the tenant master key, and stored alongside the ciphertext.
- Vault scope
- Secrets are scoped per user within a tenant. A user's vault is inaccessible to other users, including Tenant Admins. Super Admins can audit vault metadata (key names, creation dates) but never the plaintext values.
- Erasure
- Vault entries are included in both Phase 1 (immediate anonymization) and Phase 2 (physical deletion, KMS key deletion) of the NuBench two-phase erasure workflow. Requesting erasure removes the vault completely within 72 hours.
- Classification
- Secrets detected in a PHI-classified conversation are stored with PHI classification. The vault enforces the same classification ceiling as the conversation in which the secret was first captured.
- Audit trail
- Every vault write, read, and reference resolution is written to the immutable audit trail with the user identity, conversation ID, and timestamp. The audit record confirms a secret was used without recording the plaintext value.
What your team can do with it
Credential rotation without friction
When a key rotates, update it once in the vault. Every reference that uses that key name picks up the new value automatically — no prompt archaeology required.
Compliance-safe debugging
Operations teams can share connection strings and tokens in AI-assisted debugging workflows without those values persisting in conversation history, audit exports, or model context windows.