Trust

Security

How we protect your data

Private alpha note. The controls below describe Continuum's production posture. The tenant-isolation, authentication, and transcript-handling layers are operating today against the same code that will run in production; the deploy-platform controls (cloud infrastructure, audit-log retention, region selection) become live the moment Continuum is deployed for a customer. Pre-deploy, no live customer data exists.

SOC 2 posture

Our architecture is built to a SOC 2-shaped standard: tenant isolation, change management, encryption in transit and at rest, structured audit codes, and a regression test for the isolation layer that runs on every commit. We have not yet engaged a SOC 2 auditor and we have not started a Type II operating period; that work begins with the first paying customer whose procurement requires it. The controls described below are what an auditor would look at.

Isolation

Every workspace is fully isolated at the database level. Queries are scoped by workspace ID enforced in the API layer. No workspace can access another's data.

Encryption

All data is encrypted in transit (TLS 1.2+) and at rest. Transcript files uploaded for processing are stored encrypted in object storage and deleted after extraction completes.

Authentication

Authentication is handled via Zitadel, a self-hosted open-source identity provider (Apache 2.0). We support Google OAuth and email/password. Sessions use short-lived JWTs. MFA is available.

Transcript handling

Raw transcript text is stored encrypted and deleted after extraction completes. During extraction, PII-redacted transcript text is sent to a third-party LLM API (Google Gemini or OpenAI) under data-processing agreements (see the AI-assisted processing section below). The persistent record is the structured commitment data, not the source text. You control deletion of your workspace's commitment data at any time.

AI-assisted processing

Continuum uses a large language model (LLM) to extract commitments, signals, and decisions from meeting transcripts. Before any transcript text is sent to an LLM API:

—PII redaction. Phone numbers, email addresses, credit card numbers, SSNs, and IP addresses are replaced with stable redaction tokens.
—Graceful degradation. If the LLM provider blocks or errors on a transcript, extraction degrades gracefully to zero results. The meeting is still recorded; no pipeline failure, no data loss.
—In-app disclosure. AI-extracted outputs carry an explicit "AI-extracted · verify before acting" disclaimer throughout the product.

Prompt-injection detection and a harmful-content pre-ingestion gate are in active development and will be documented here when shipped.

Infrastructure

Hosted on Google Cloud Platform. Stateless services run on Cloud Run (auto-scaling, ephemeral). Stateful services (database, auth) run on isolated GCE instances with private networking. No public database endpoints.

Audit log & observability

Tamper-evident audit logs and end-to-end observability run on our home-grown fasten.sh: append-only at storage, structured event codes, actor-enriched. A workspace admin reviews the same audit trail an external SOC 2 auditor would. Coverage includes ingestion, extraction, commitment changes, and membership changes. Audit event codes are defined and categorised; the live emission wiring and retention infrastructure ship with the production deploy.

Security roadmap

Security controls ship in four phases tied to the product's delivery milestones. Here is what's in place and what's coming.

—Pre-launch (operational now). Four-layer tenant isolation — schema-per-tenant, row-level security (RLS), unprivileged DB role, trap schema — is fully deployed and regression-tested. Auth bypass is refused at boot in production. OpenAPI docs are hidden outside debug mode. CORS is locked to known origins. PII is redacted on every LLM path. Extraction failures degrade to zero results; no pipeline failure, no data loss.
—v1 hardening (within 30 days of first live customer). JWT (JSON Web Token) token revocation via blocklist, logout propagation to the identity provider, and user-deactivation endpoint ship together. Role-based access control (RBAC) is centralised in a FastAPI dependency and verified by CI lint — no endpoint can ship without an explicit authorisation intent. Prompt-injection phrase filtering and a harmful-content pre-ingestion gate land on the sanitizer layer. Audit events move from code stubs to live emission.
—Enterprise gate (before any enterprise contract). MFA (Multi-Factor Authentication) enforcement becomes policy-driven per workspace. Meetings can be marked Legal/HR/Privileged: no AI extraction is performed, no transcript text sent to any provider. All operator database access is proxied and every break-glass action is logged. Self-serve workspace data deletion ships. The in-product audit log is visible to workspace admins.
—SOC 2 posture (6–12 months). Schema migration audit trail (Alembic), 3-year structured audit log retention, GDPR (General Data Protection Regulation) right-to-erasure tooling, and KMS (Key Management Service) per-tenant encryption are the long-term targets. SOC 2 Type I assessment follows once operational controls are stable.

Our detailed security posture, gap tracker, and phase delivery timeline are available to enterprise buyers under NDA on request. hello@continuumstate.io.

Vulnerability disclosure

If you discover a security vulnerability, please report it to hello@continuumstate.io. We respond within 48 hours and will not pursue legal action against good-faith researchers.