Security & data sovereignty
Your client work is yours.
Most legal AI tools take ownership of your documents the moment you upload them. They chunk your matter files into a vendor-hosted vector database, feed them to a third-party model, and reserve the right to “improve their service” with your data. Jubal is engineered the other direction. Three promises follow — each made structurally enforceable by the architecture below.
We never train on your work.
Not Jubal. Not any of our model providers. Not anyone. Your matter files, the prompts you send, and the documents Jubal generates for you are governed by no-training API posture across every model provider Jubal uses, plus our own contractual commitments to those providers. Jubal does not have a separate “use customer data to improve our product” pathway and we will not add one. If we ever change this, it will require an opt-in from your firm — and the default will always be no.
We never share your work.
Not with our partner firm. Not with academic researchers. Not with other Jubal customers. Not aggregated, not anonymized, not summarized, not statistically. The only circumstance in which Jubal would disclose customer data is a properly served legal process targeting your firm — at which point we would notify you immediately to the extent the law allows. Beyond that, your work moves only between the seats your firm authorizes and the specific frontier-model API calls needed to answer your prompts.
Your matters stay yours.
Each attorney runs Jubal against a private index on their own machine. For multi-attorney matters that span offices, Jubal supports secure firm-scoped sync — content is encrypted in transit and at rest, and access is scoped to the seats your firm authorizes. When you ask Jubal a question, only the small slices of text needed to answer that specific question are sent to a frontier-model API, behind a no-training agreement. Jubal has no use for your matters beyond answering the prompt you sent: we don't keep them for training, we don't keep them for evaluation, and your matters are never used to improve our product or anyone else's.
How the promises are structurally enforced
Local indexing, firm-scoped sync
Each attorney runs Jubal against a private index on their own machine. Each matter folder is scanned with a fast, local search index (ripgrep + manifest ranking). When you ask a question, Jubal's client decides locally which slices of which files are likely relevant and sends only those slices to the frontier model — no chunking-into-a-vendor-vector-database step, no upload of your full corpus to a third-party service. For multi-attorney matters that span offices, Jubal supports secure firm-scoped sync: content moves only between the seats your firm authorizes, encrypted in transit and at rest. The boundary is the firm, not a single machine.
Model posture
Jubal routes each task to the best frontier model for it, across every major provider — Anthropic, OpenAI, Google. For stronger results, Jubal sometimes blends outputs from more than one provider. Every provider Jubal uses is called under the no-training API posture each provider offers to API customers, plus our own contractual commitments. Your prompts and the slices of text we send with them are not used to train, tune, or evaluate any model — ours, theirs, or anyone else's.
The partner-firm data wall
Jubal's partner firm is a separate corporate entity from Jubal. The partner firm's attorneys are co-designers of the product; they are not employees of Jubal and they do not have access to Jubal's production environment. There is no shared infrastructure between Jubal's production environment and the partner firm's practice environment. The participation agreement between the two entities forbids cross-access. When the founding cohort opens, you can request a copy of the relevant section of that agreement.
Encryption at rest and in transit
Connections between the desktop client and Jubal's services are TLS 1.3. The local index on each attorney's machine can be configured to live inside an encrypted volume (FileVault on macOS, BitLocker on Windows). Firm-scoped sync content is encrypted at rest in Jubal's storage layer and in transit between firm seats. Backups, where your firm asks for them, are encrypted with keys managed by your firm.
Audit logs and matter walls
Jubal records who at your firm asked what, when, and against which matter — locally, in your firm's own audit log. Per-matter walls are first-class: you can scope a Jubal session to a single matter and prevent it from drawing on documents outside that scope. The wall is enforced at the index layer, not just in the prompt.
The partner-firm data question, answered again
“Wait — your partner firm? Does that mean my matter data is going to a competitor of mine?”
No. Our partner firm is a co-designer and investor. They helped build Jubal. They do not have access to Jubal customer data — yours or anyone else's. The data wall is enforced architecturally (separate corp entity, separate production environment, no cross-access) and contractually (their participation agreement explicitly forbids it). When the founding cohort opens, you can request a copy of the relevant section of the agreement.
Compliance posture
Honest about what is and isn't certified yet:
- SOC 2 Type II — planned, not yet held. Targeted for completion ahead of general availability.
- HIPAA-aware posture for healthcare-adjacent practice (PI, medical malpractice, employment) — planned.
- Single sign-on (Okta, Google) — planned.
- State-bar ethics opinion roundup— planned public reference page covering the major published opinions and how Jubal's architecture maps to them.
Contact for IT and security teams
Send your security questionnaire to security@jubal.law. We'll respond inside three business days.
Questions about Jubal's architecture from a non-security angle? Read more on the FAQ, or write to hello@jubal.law.