VERITY Security Model

Data in transit — TLS#

Every network hop is encrypted with TLS.

• Application traffic. Production traffic is served over HTTPS only. TLS is terminated at the edge, and any HTTP request is redirected to HTTPS.
• Database connection. The database connection requires SSL in any non-local environment. This is enforced at startup — if the TLS requirement is missing, the application refuses to boot.
• AI provider calls. All calls to Anthropic (Claude) and Deepgram (transcription) travel over HTTPS via the providers' official SDKs.
• File uploads. Uploads and downloads of client files are HTTPS-only.

Data at rest#

• Database. Client data is stored in PostgreSQL hosted on Neon, which encrypts stored data at rest by default (AES-256 on the underlying AWS-managed volumes).
• Uploaded files. Uploaded files are stored with private access. Unauthenticated attempts to reach a private file's URL return a 403. Files are served only through an authenticated download route that verifies ownership against the signed-in professional's session before streaming the file back. Files uploaded before the private-storage cutover are being migrated; until that migration completes, those earlier files may remain reachable at their original URLs.

Application secrets#

• Session signing key. Professional sessions are signed with a server-side secret that is required at startup. Rotating it invalidates existing sessions.
• API keys & tokens. Provider API keys (Anthropic, Deepgram) and the file-storage token are stored as server environment variables. They are never shipped to the browser and are consumed only in server-side code paths.

Abuse & rate limiting#

Requests to VERITY's API are protected by a distributed, sliding-window rate limiter (Upstash Redis). If that limiter is not configured in a production environment, the application fails closed — it refuses to run unprotected rather than serve traffic without limits. During a transient Redis outage at runtime, account registration fails closed, while in-progress client and professional requests are deliberately allowed through so an outage never strands a client mid-conversation — a documented availability trade-off, not an oversight.

What the “Encrypted” indicator does & does not claim#

Audit trail#

Raw client inputs — messages, audio transcripts, and uploaded documents — are append-only and are never hard-deleted. A soft-delete is available at both the intake and professional level, but the underlying record history is preserved for forensic auditability.

Known limitations#

We would rather tell you where we are not yet than imply a maturity we have not reached:

• No centralized observability (log drain or error monitoring) yet. Production errors surface only in the hosting platform's runtime logs.
• No formal penetration test has been performed.
• No user-facing privacy policy, Terms of Service, or self-service data-deletion route yet.
• Prompt injection: client content is fenced as untrusted data in every AI prompt, every extracted fact's citation quote is verified as a verbatim substring of what the client actually said (unverifiable quotes are downgraded, never trusted), and a contradiction can only be resolved by the client's verbatim clarifying words. We now run an adversarial injection test suite against this path — attacks that try to make the assistant declare the intake complete, reveal its instructions, switch into giving legal advice, break out of the untrusted-content fence, or plant fabricated facts and citations, in both English and Spanish. In our most recent run the assistant resisted every case, and the citation-verification layer rejected every fabricated quote. Stated plainly: this is mitigation, not proof — the suite is small and not exhaustive, injected text can still color how the AI reads the data around it, and no formal third-party penetration test has been performed.
• Rate limiting fails open (except registration) during a transient Redis outage, and files uploaded before the private-storage cutover remain at their original URLs until the migration completes — both described above.

Questions#

Security questions get direct answers, not stalling. Reach out to Ryan Gruzen.