Security & trust

Your data runs the paper. It never leaves your control.

Paperbeam reads your company’s most sensitive tools so everyone can read the news. We take that seriously, and we’d rather show the work than ask for trust.

SOC 2 Type I: in progressZero-retention AINo model trainingRow-level security

The short version

Six commitments, kept in code.

Each of these is enforced by the system, not by a policy document. Where something is still on the roadmap, we say so.

Encryption

Sealed in transit and at rest.

Everything we store is encrypted, and the secrets that open your tools get their own envelope.

TLS 1.2+ for every connection, in and out
AES-256 encryption at rest for databases and files
Per-workspace keys via KMS envelope encryption for tokens and API keys
Secrets decrypted only at call time, and never logged

Tenant isolation

Your paper never prints someone else’s news.

Isolation is enforced by the database itself, not by remembering to add a filter.

Every table carries a workspace_id
Postgres row-level security on every tenant table
Per-workspace storage prefixes with server-side encryption
Signed file URLs that expire in five minutes

Retention you control

We keep what the paper needs. Then we let go.

Raw source content is short-lived. What remains are the facts and excerpts behind each story.

Raw source content held for a short window, then deleted
Derived facts and short excerpts kept to power the archive
Configurable retention per workspace and per source
Data deleted after cancellation, per our published policy

AI data use

Your data is never training data.

Models read your sources to write the paper. That is all they do with them.

Zero-data-retention model endpoints by default
Never used to train any model, ours or anyone’s
Source content is treated as data, never as instructions
Models never get write tools in your systems

Access controls

The right people, and only them.

Readers and admins sign in with the identity provider you already trust.

Google and Microsoft SSO on every plan
SAML SSO and SCIM provisioning on Business
Audit log of admin and configuration changes
Read-only source scopes by default, each one justified

Your keys, your models

Bring your own model key.

Business workspaces can route AI calls through their own provider account.

BYOK for Anthropic, OpenAI and Google
A ZDR-only setting that filters out any provider without zero retention
Token usage shown per edition
Keys stored with KMS envelope encryption

Verification

How a fact gets into the paper.

A newspaper is only as good as its fact-checking. Ours runs on every sentence, every morning, before anyone reads a word.

01
Collected
Read-only pulls from the sources you connect. Every record keeps a link back to where it came from.
02
Computed
Numbers are calculated by code from the records. Models never do arithmetic for the paper.
03
Checked
Quotes are matched verbatim against the transcript or message. No match, no quote.
04
Cited
Every sentence cites a source. A claim without support is cut before it reaches the page.
05
Approved
Editions wait in the galley for your approval, or run on their own once you trust them.

The verification gate is strict on purpose: number accuracy and quote accuracy must be 100% before any model is allowed to write for a paper.

Compliance

Honest about where we are.

We are in private beta. Here is exactly what is done, what is underway and what comes next.

SOC 2 Type I: In progressNot yet certified. We are working toward a Type I report before public launch, with Type II to follow. The report will be available under NDA.
Penetration test: PlannedAn independent pen test before public launch, then annually.
GDPR and DPA: AvailableOur Data Processing Addendum includes Standard Contractual Clauses and a published subprocessor list.
Vulnerability disclosure: OpenReport issues to security@paperbeam.ai. We publish a security.txt and respond to good-faith research.

Subprocessors

Who else touches the data.

A short list, chosen carefully. We notify customers before adding anyone new.

Full list and change notices

Vendor	Purpose	Data	Location
Vercel	Web hosting and edge delivery	Request metadata; content in transit	United States (global edge)
Neon	Postgres database	Account data, derived facts, editions	United States
Cloudflare R2	File storage	Evidence packs, audio, PDFs, images	United States
Railway	Background workers and page rendering	Content in processing; not stored	United States
Anthropic	AI model provider (zero data retention)	Source excerpts in prompts; not retained	United States
OpenAI	Text-to-speech and fallback models	Edition text; source excerpts on fallback	United States
ElevenLabs	Audio edition voices	Edition script text	United States
Resend	Email delivery	Recipient addresses, edition emails	United States
Stripe	Billing and payments	Billing contact, payment details	United States
Nango	Integration authorization and sync	OAuth tokens (encrypted), synced records	United States
PostHog	Product analytics	Usage events on our site and app; no edition content	United States
Sentry	Error monitoring	Error traces, scrubbed of content	United States

trust.paperbeam.ai (coming soon): live controls, policies, uptime history and report requests in one place.

Request a security review

Every sentence has a source. Including this page.

Questions about security? Write to security@paperbeam.ai and a person will answer.

Start free: 10 editions on us See your paper