Security at B3IQ. Data boundaries, made clear.

We do not write inference requests or responses to storage. A request is dispatched in memory to the machine that serves it and discarded when the request completes.

We do not train on your data and we do not sell it.

What we do record, per request, is the metadata needed to route, measure usage, and prevent abuse: a request identifier, timestamp, model, serving machine, token counts, latency, and status. That metadata contains no prompt or completion text, and it is deleted after 90 days.

Features you explicitly opt into may store more, and say so where you turn them on. For example, playground conversations you save are stored so you can return to them, and are deleted when you delete the session. Encrypted private jobs are held only as ciphertext we cannot decrypt, and purged after about 30 days.

You own the machine. B3IQ builds it in Oregon, hosts it in the United States, and runs it as your private AI inference endpoint.

B3IQ routes each customer’s requests to that customer’s own machines. You are not sharing an inference pool with other tenants, and a key can be narrowed further to a named set of machines.

Traffic to B3IQ is protected with TLS in transit. Data at rest lives on our infrastructure provider's encrypted storage. Credentials are never stored in a form we can read back: API keys, access tokens, and recovery codes are stored as one-way hashes and shown to you exactly once, and account passwords are hashed with bcrypt.

End-to-end encrypted inference is available. On this path your client encrypts the request to the serving machine's public key, and B3IQ relays the ciphertext without holding a decryption key: the platform is cryptographically excluded from your content. The machine's encryption key is short-lived and is signed by that machine's own identity key, which your client verifies before encrypting.

It protects your content from B3IQ's own systems and from our public surfaces. It does not hide content from whoever operates the machine doing the work. When that machine is yours, that is you.

Staff access to production runs through a single-sign-on gateway restricted to our corporate domain, and administrative roles are separated so that read-only, operational, and money-moving actions require different authority.

Every administrative request is recorded in a durable audit log with the actor, the action, the result, and a hashed source address. High-risk actions additionally require a written reason, and the most consequential ones require a typed confirmation before they will execute. Where staff need to diagnose a problem on a machine, they use a time-limited, revocable credential scoped to read-only diagnostics, not a standing login.

The service enforces request size limits, token limits, and per-key, per-account, and per-machine rate limits, and supports per-account spend caps and model restrictions. Operators can pause a class of traffic, an account, or a single machine immediately, without a software release, if something needs to be stopped.

Machines that stop responding, fail readiness checks, or report an unhealthy queue are removed from routing automatically.

Live availability is public at status.b3iq.org. It covers the inference gateway, the model catalog API, this website, and the documentation, checked from multiple regions, and it is where we post incidents. It runs outside our own infrastructure.

Changes to production go through reviewed pull requests. Each one runs static security analysis, dependency scanning and secret scanning. Platform changes also run type checks, tests and a billing and settlement suite against a real database.

Software we distribute is checksummed and its release manifest is cryptographically signed, so the binary you install can be verified against what we published.

We use a small number of vendors to run the service: our cloud infrastructure and content-delivery provider, payment processors for card and crypto purchases, a transactional email provider, analytics providers, a code-signing service, our source-control and CI provider, and a model artifact repository.

Current customers and prospective customers under assessment can request the itemized subprocessor register, naming each vendor and the data it receives, from the contact below.

• SOC 2 Type IIUnderway with A-LIGN, report expected in 2027
• HECVAT (full)Completed
• HIPAA Business Associate AgreementWe sign it
• Written security policy setAvailable on request
• Security awareness trainingRequired for everyone with system access
• FedRAMP 20x readiness assessmentNext, with A-LIGN

The policy set covers information security, access control, change and patch management, incident response, business continuity, data handling, key management, privacy and AI risk.

If your review needs something not listed here, tell us.

Email contact@npclabs.org with the subject line SECURITY, or see /.well-known/security.txt. We aim to acknowledge reports within three business days.

We will not pursue legal action for good-faith research that respects customer data, avoids degrading the service for others, and gives us a reasonable opportunity to fix the issue before disclosure. Please do not access, modify, or retain data that is not yours, and please do not run automated scanning that affects availability.

Security and privacy correspondence goes to the address below.