SECURITY

Security you can check, not just trust.

A private AI server is only as safe as the way it's run. Here's how we set up and operate yours.

Plain talk first

Designed around your obligations

Every deployment is configured for the security and privacy requirements your organization works under. For organizations that handle health information, the HIPAA edition is HIPAA-aligned, with locked safeguards. Compliance depends on your whole program, and we document our part so your reviewers can check it.

Built to a high bar

Encryption, multi-factor sign-in, network segmentation, restore planning and regular testing are part of the design, not add-ons. The HIPAA edition is built to the proposed HIPAA Security Rule updates, which aren't final yet.

On-site isn't automatically safer

One less internet-facing service holding your data shrinks the blast radius. But a neglected server in a closet is worse than a well-run cloud. That's why monitoring, patching and testing are part of every engagement.

Your data stays on your server

Documents, chats, search indexes, transcripts and audit logs are stored on encrypted drives on a server in your building that you own. Backups go to storage you control, encrypted with a key you hold.

  • HIPAA edition: never sends prompts or health information to a cloud AI service.
  • Standard edition: cloud AI is off unless you opt in under your own contract.
  • What reaches us: our update and monitoring systems are designed to hold none of your data. Your server sends us only health metrics and error codes, filtered against an allow-list, with free text removed.
  • Air-gapped option: no internet connection at all, with updates delivered on signed USB media.

Your server

  • Documents
  • Chats
  • Search index
  • Audit log

Saddle Tree updates & monitoring

The controls, edition by edition

In the HIPAA edition, these are locked settings. A customization can't turn them off.

ControlHIPAA editionStandard edition
Encryption at restFull disk, unlocked by the TPM chip plus a network key or PIN. A stolen server won't unlock off siteFull disk, unlocked by the TPM chip
Encryption in transitTLS for every browser connection, and mutual TLS between services inside the serverTLS for every browser connection
Sign-inSingle sign-on with your accounts. Multi-factor required and checked at every login. No shared accounts. Logoff after 15 minutes idleSingle sign-on with your accounts. Multi-factor strongly recommended
AccessRoles control models, documents and workflows. Each workflow uses only the data fields its job needs. Quarterly access-review reportRoles control models, documents and workflows
Departing staffLose access at the next directory syncLose access at the next directory sync
Audit logEvery prompt, response, document retrieved, workflow action, admin change and support session. Tamper-evident and append-only. Reading the log is logged tooSign-in, admin and configuration events. Prompt logging optional
BackupsNightly, encrypted, plus an offline copy on storage you control. Quarterly restore test with written evidence, and a documented restore planNightly, encrypted, to storage you control
TestingVulnerability scans every six months and an annual third-party penetration testAnnual vulnerability scan
End of contractData returned or destroyed, including cryptographic erasure, with a certificate of destructionData returned or destroyed

Remote support is off by default; we connect only when you approve each session.

Your admin approves each time-limited session in the dashboard. Every session uses multi-factor sign-in and is recorded in your audit log. In the HIPAA edition, sessions are short, come only from our company-managed, encrypted laptops, and no files are copied off your server. During an approved session our team may see your data, which is why access is approval-only and every session is logged.

Updates that check themselves

  1. Step 01: Signed

    Every update is cryptographically signed. Your server checks the signature before installing anything, and pulls updates over an outbound-only encrypted connection. We never need an open port into your network.

  2. Step 02: Tested

    Each update runs on our lab system and an early-release site first. Model changes happen only after they pass quality tests on your own examples.

  3. Step 03: Reversible

    Your server backs up before updating, runs health checks after, and rolls itself back automatically if anything fails.

Routine updates are monthly. Critical security patches are expedited. You approve maintenance windows, and you can put updates on hold.

On its own network, in a locked room

  • Its own network segment, with no inbound connections from the internet
  • Your data and management traffic on separate internal networks
  • The component that reports health to us can't read the data drives
  • Battery backup with automatic safe shutdown
  • A locked, ventilated closet, an asset tag and recorded serial numbers
  • Burn-in and imaging in our lab, with a documented chain of custody

Drafts, not decisions

Every output is a draft until a professional on your team approves it. Nothing is written directly into your core systems. In the HIPAA edition, outputs are labelled "AI draft: requires professional review," and clinical-decision uses are switched off. Our software doesn't diagnose, give legal advice or give tax advice.

Have a security questionnaire? Send it over.

We'll answer it plainly, and tell you where the answer is "not yet."