Security & Compliance

Architecture & Data Flow

Sevrel's checked-in deployment model uses a static Next.js frontend, a FastAPI application, PostgreSQL and Redis, configured document sources, and external AI or integration providers. Anthropic Claude is the default LLM implementation; optional providers and integrations vary by organization and deployment.

Data Encryption

The repository also contains a decrypt-fallback and resumable re-encryption workflow for controlled application-key retirement. Key custody and execution of a rotation require operational proof.

Tenant Isolation & Authorization

Sevrel combines several defenses rather than relying on one universal query invariant:

• Organization and user identifiers on customer-data models
• Explicit tenant filters in routers and services
• PostgreSQL row-level security as a database boundary
• Fail-closed tenant session context
• Route, RLS-coverage, IDOR, and real-Postgres cross-tenant tests
• Three base roles, organization-defined custom roles, and named permissions for privileged operations

The production database role, grants, and enabled/FORCE RLS policy state must be verified from the live database. Sidebar visibility is not an authorization control.

Authentication & Sessions

• Microsoft Entra ID and Google OAuth paths validate provider identity assertions server-side.
• Password/TOTP accounts are also implemented; passwords are bcrypt-hashed rather than reversibly stored.
• Session cookies are HttpOnly and SameSite=Lax, with Secure required by production configuration.
• Active server-side session rows support individual and account-wide revocation.
• High-impact email and calendar writes require short-lived, single-use confirmation tied to the requesting user.

Organization MFA, conditional-access, and account-lifecycle policy remain controlled by the customer's configured identity provider and require tenant-specific evidence.

Audit Logging & Observability

Selected security and product events are written with user, organization, action, timestamp, and correlation context. A database trigger rejects audit-row update and deletion except for a narrowly constrained erasure-pseudonymization path.

This is append-only database enforcement, not cryptographic immutability or WORM storage, and it does not mean every customer-data read is logged.

The web, worker, and scheduler entry points share a recursive fail-closed Sentry scrubber. Sentry is live for backend services, operator-verified on July 26, 2026, with a privacy-safe smoke receipt, a production-scoped high-priority email alert rule, and the monitoring plan's published 30-day event lookback. Alert-mailbox receipt and formal 24x7 response coverage are not verified.

Backup & Recovery

The repository contains age-encrypted logical-backup tooling and an isolated restore verifier. A production plaintext-path dump-and-restore drill was recorded on July 14, 2026; it predates the current encrypted, application-role-aware verification path.

Repository automation does not schedule logical backups. On July 28, 2026 an operator observed scheduled daily, weekly, and monthly Railway volume backups with recent artifacts in the signed-in production dashboard; point-in-time recovery is not enabled or offered for the current custom-image service, and no restore was executed during that check. Those observations do not prove database-consistent restoreability, contractual retention, vendor assurance, RPO, or RTO. Sevrel does not publish a guaranteed RTO or RPO without current operational and contractual evidence.

Responsible AI

• Retrieval and citations help users inspect the evidence used for an answer, but do not eliminate hallucinations.
• Sanitizers reduce recognized PII and secret exposure at defined high-risk boundaries; they are not a universal detection guarantee.
• Sevrel does not use customer content to train its own models. External-provider training and retention are governed by current provider terms and the customer's executed agreement.
• Anthropic Zero Data Retention is not yet in place. Under Anthropic's standard API terms, API request inputs and outputs are retained for up to 30 days for abuse monitoring — longer only if flagged for Trust & Safety review — and are not used by default to train Anthropic's models. Sevrel will not assert Zero Data Retention without written account-level confirmation from Anthropic.

Compliance Status

Sevrel has not completed a SOC 2 Type I or Type II audit, has no active observation window, and has not completed an independent penetration test. A SOC 2 readiness program is in progress: a control registry, a SOC 2-oriented self-assessment, and an evidence index are maintained internally.

Data retention, regional processing, subprocessor terms, incident notification, and deletion commitments are governed by executed agreements and live evidence, not by repository-compatible code alone.