This page is written to be a straight answer for lawyers, security reviewers, and clients doing due diligence on Wythe. Where we are early-stage, we say so plainly. It is not legal advice.
The short version
- We do not train on your data — and our AI providers are contractually barred from training on it either.
- Your matter content is visible only to your organization, enforced by the database, not just by application code.
- Your search queries are never stored.
- Delete a matter and it is permanently purged after a 7-day grace window; backups roll off within 35 days.
- Built to the SOC 2 Trust Services Criteria on independently audited infrastructure — with an honest account below of what we have and what is still on the roadmap.
On this page: Your data & who can see it · AI & model training · Third parties · Retention & deletion · Security engineering · Confidentiality & privilege · For political & campaign clients · Compliance status · Reporting a vulnerability
Your data and who can see it
Who can see my matter content?
Only authenticated members of your organization. Every research and AI feature requires sign-in, and that is enforced by a build-time check that fails our release if any API route is missing its authentication guard. Your matter content is scoped to your organization by PostgreSQL row-level security — a database-level control, so isolation does not depend on application code getting every query right. See also who at Wythe can access data.
What does Wythe log about my research activity?
We do not store your search terms, your filter values, or which documents you open. Our usage telemetry is content-free by construction: a server-side allowlist limits it to counts, timings, and feature/operator types — never the values — so content cannot leak into logs by accident. We run no third-party analytics (no Google Analytics, no Mixpanel, no advertising pixels). If a page fails to load, a diagnostic report that would include content is sent only with your explicit, per-incident consent, which you control in the app.
Where is my data stored?
In the United States. Our database and authentication (Supabase), object storage and AI worker (DigitalOcean), and application hosting (Vercel) all run in U.S. regions. See Subprocessors for the full list.
AI and model training
Do you train AI models on my data?
No. Wythe does not train, fine-tune, or evaluate models on your matter content, your queries, or the outputs generated for you.
Do your AI providers train on my data?
No. Wythe AI runs on Anthropic'sClaude models through the commercial API, and Anthropic's Commercial Terms of Service state that Anthropic does not train its models on API customer content. Anthropic maintains its own independent audits — SOC 2 Type II, ISO 27001, and ISO 42001 (the AI-management-system standard). Content is retained only transiently to serve your request; we are pursuing formal zero-data-retention terms with our providers and will update this page when they are in place. (Providers may retain content that their automated trust-and-safety systems flag, or that law requires them to keep.)
What exactly is sent to the AI provider when I use Wythe AI?
Your prompt, the relevant matter files, and Wythe's instructions — over an encrypted connection, to produce the research or draft you asked for. Uploaded scanned documents may also be sent to Mistral for OCR conversion. Nothing is sent to any AI provider when you use plain corpus search. Uploaded content is passed to the model as data, never as instructions.
Does Wythe AI search the web, and what leaves the system?
Web and case-law search are off by default. When you turn them on, the model composes search queries from your conversation and sends them to Anthropic's web-search tool (results provided by Brave Search, an Anthropic subprocessor) and to CourtListener (the Free Law Project's public case-law database). Your uploaded documents are never sent to these services — only the model-composed queries, which may reflect the substance of what you are researching. That is why the feature is opt-in.
How do you defend against prompt injection?
We design to bound the blast radius rather than promise immunity. The AI worker has a small, fixed set of tools, and the only one that writes anything is a memory-note tool. Uploaded content is delimiter-escaped and treated as data. Most importantly, the worker is confined at the database layer for each run — even a fully hijacked run can only read and write its own matter. That isolation is exercised by automated tests, including prompt-injection scenarios, that must pass before we ship.
Third parties
Which third parties process my data?
The providers that operate the Service are listed, with what each receives and its independent audits, on our dedicated Subprocessors page. In summary: your matter content is handled only by Supabase (database), DigitalOcean (storage), Anthropic (AI drafting/research), and Mistral (OCR); billing is handled by Stripe (full card numbers never touch Wythe); and the public FEC corpus is processed by additional model and embedding providers that never receive your matter content. Every infrastructure provider that touches customer data is independently SOC 2 Type II audited.
Do you use analytics or advertising trackers?
No. We use strictly-necessary cookies for your session only — no advertising technology, no cross-site tracking, and no third-party analytics of any kind.
Retention and deletion
What happens when I delete a matter?
It becomes inaccessible in the product immediately, then enters a 7-day grace window (protection against mistakes). After that it is hard-purged — object-storage originals first, then every database row — leaving only a content-free audit record. Residual copies in our database provider's encrypted disaster-recovery backups roll off within 35 days and are never accessible through the product.
Do you keep my search queries or AI chat history?
Corpus search queries are never stored — we deliberately removed the table that would have held them. Your Wythe AI chats and drafts are stored, because they are your work product: they live inside your matter, scoped to your organization, and are purged with the matter as above.
How do I delete my account?
Email info@wytheintelligence.com and we will delete your account and its content within 30 days, subject to any legal retention obligations. Self-serve deletion is on the roadmap.
Security engineering & operations
Encryption
Data is encrypted in transit with TLS 1.2+ (with HTTP Strict Transport Security) and at rest with AES-256 through Supabase and DigitalOcean.
Tenant isolation
Row-level security scopes every user-facing table to the owning organization, and the AI worker is additionally confined per-run at the database layer (see prompt injection). Both are covered by automated tests in our release pipeline.
How is the codebase itself secured?
Every change reaches production through a pull request with automated checks: type checking, linting, tests, a secret scanner over the full git history, and the authentication guard described above. No one pushes directly to production, and secrets live only in environment configuration, never in source control.
Do you support MFA or SSO?
You sign in with an email and password (hashed by our authentication provider; we never see it in plaintext) or an emailed one-time link. Two-factor authentication (TOTP) is available— turn it on under Security in your account settings, and you'll be asked for a code from your authenticator app at each sign-in. SSO/SAML is planned for enterprise customers; we'll state it as available here only once it is.
Incident response and breach notification
We run independent uptime monitoring, content-free client-side error telemetry with consent-gated diagnostics, and a daily automated operations review. Our incident playbook is to rotate affected credentials, force-expire sessions, review access and provider logs, and notify affected customers by email without undue delay. No system is perfectly secure; keep your credentials confidential and tell us promptly of any suspected unauthorized access.
How do I report a vulnerability?
Email security@wytheintelligence.com (or info@wytheintelligence.com). We welcome good-faith security research and will acknowledge your report promptly. Please give us a reasonable chance to remediate before any public disclosure; we will not pursue researchers who act in good faith and avoid privacy violations, data destruction, and service disruption. Our machine-readable contact is at /.well-known/security.txt.
Confidentiality, privilege & professional responsibility
Wythe Intelligence LLC is not a law firm and this page is not legal advice. The judgment about what you may put into any third-party tool is yours to make. What we can do is give you the facts your analysis needs.
Does putting client material in Wythe risk waiving privilege?
That is a professional-responsibility judgment for you, informed by rules like ABA Model Rule 1.6 and ABA Formal Opinions 477R and 512 on reasonable efforts to protect client information and on generative AI. We do not and cannot decide it for you or assert that privilege is preserved. This page exists to be your due-diligence record for that analysis. The facts it depends on:
- No training on your content, by us or our AI providers (AI & model training).
- Encryption in transit and at rest (encryption).
- Organization-scoped isolation enforced in the database (who can see my content).
- Customer-controlled deletion (retention & deletion).
- United States hosting (where data is stored).
- Independently audited subprocessors (subprocessor list).
- A DPA on request (compliance status).
Our Terms of Service (Section 8) also record that you own your Matter Content, that AI output must be verified before filing, and that you should not upload material you are not permitted to store with a third-party processor.
Who at Wythe can access my data?
We are a small team and direct about this: administrative access to production systems exists and is limited to the founding team. It is used only to operate and support the Service — for example, to investigate an incident or a support request you raise. We do not browse customer matters. Formal, logged access reviews are part of our SOC 2 program build-out. We do not claim that no employee can access data — that would be false for any operated service, and we would rather tell you the real boundary.
Can anyone tell what my firm is researching?
The research corpus is public record, so reading it discloses nothing to anyone. Your queries are not stored, our telemetry is content-free, and your matter content is organization-scoped. There is no analytics or advertising party in our stack that could observe your research.
For political & campaign clients
We handle donor data and opposition research. Who can see it?
No one outside your organization. The same protections apply as to any matter content: nothing about your reading of the public corpus is recorded with content, your queries are not stored, and — notably for this space — there is no ad-tech or third-party analytics anywhere in our stack.
What happens to uploaded documents like donor lists or research files?
They get the same protections as all matter content: no model training, database-enforced isolation, and hard deletion on your schedule. In the interest of full disclosure, scanned PDFs pass through Mistral's OCR API for text extraction (see Subprocessors). As our Terms note, the decision about what to upload is yours.
Compliance status & working with us
Are you SOC 2 certified?
No — and strictly speaking, no vendor is SOC 2 “certified”: SOC 2 is an independent auditor's attestation, not a certificate anyone issues. Wythe's controls are built to the SOC 2 Trust Services Criteria, and a Type II examination is on our roadmap. Meanwhile, every subprocessor that touches customer data holds its own independent audit (see the subprocessor list).
Do you offer a DPA?
Yes — we can sign a Data Processing Addendum covering data handling, the no-training commitment, subprocessors, and breach notification. Email info@wytheintelligence.com to request our current terms.
Will you complete our security questionnaire?
Yes, and we turn them around quickly. Email info@wytheintelligence.com.
What is on the security roadmap?
These are planned, not yet in place:
- SSO/SAML for enterprise.
- Self-serve account deletion.
- Formal, logged access reviews.
- An independent penetration test.
- A SOC 2 Type II examination.
Questions, or a security review to run? Email info@wytheintelligence.com.