Security

Access is scoped, not filtered

Every read of a whiteboard, canvas, file or meeting resolves the caller’s organization and department first, and queries are constrained to what that person can reach. The scoping is in the data access layer rather than in the interface, so a surface added later inherits it instead of having to remember it. Views that look across whiteboards, like the Continuity Index and Contradiction Watch, only ever show whiteboards you could already open.

An Executive sees across their organization’s departments. A member sees their own department, plus anything explicitly shared with them. Only a whiteboard’s owner can change it. Nothing leaves your organization unless someone in it publishes a whiteboard as a public link — which is deliberate, reversible, and shows a clean record rather than the workspace. A link can be set to expire, and one control on your Account page turns off every link you have at once.

Joining an organization

People join through an invite link that works once and expires after seven days, and only when a seat is free. We store only a scrambled form of each link, so a copy of our database could not be turned into working invites.

Support cannot read your workspace

There is no back door for support to open a customer’s whiteboards. Access requires consent you grant from your own account settings, lasts at most two hours once granted, and can be revoked at any time. Every request and every decision is written to our audit trail and to your organization’s own activity log. This is the control we would most want to see as a buyer, so it is the one we built first.

What the AI models see

Text you write on a whiteboard, files you upload and meeting transcripts are sent to Anthropic and OpenAI to pick out decisions, risks and questions and how they connect. A Scratchpad draft is sent to OpenAI when you pause, so Echo can say what it is close to, and a note you speak is streamed there as you talk, so the words appear on screen, then sent once more to be written down. Once you finish speaking, the note’s text goes to Anthropic so Echo can ask one question about it, which you can switch off. Neither the draft nor the audio is stored by us. They process it for us under their business terms, which do not allow training on it, and we do not train models on it either. Under those terms each provider may keep what it received for up to 30 days to detect abuse, then deletes it; we have not negotiated zero-retention terms with either.

Every item picked out shows the exact words it came from, and a quote is kept only if it really appears in what was written, so the record cannot put words in anyone’s mouth. Anything the assistant suggests on its own is labelled as a suggestion, not presented as something a person said.

What you can check and take away

• Activity log. Your organization’s owners and admins see who joined or left, invites, sharing, departments, branding, support access, settled conflicts and exports, and can download the whole log as a spreadsheet.
• Your data. Anyone can download everything on the whiteboards they own, as one file, from their Account page.
• Your organization’s data. Owners and admins can download every whiteboard shared with the organization or filed in one of its departments, with its members and activity log.
• Closing an account. When an owner closes the organization’s account, everyone has 30 days to download their data. After that it is permanently deleted, as the Privacy page describes.
• Deleting your own account. Anyone can ask for their account to be deleted, with the same 30 days to download and change their mind.

Administrative surfaces

• The internal admin area is gated by both a database role and an email allowlist. Either alone is not enough.
• Administrative sessions require a second factor (TOTP) every time, enrolled per-account. An administrator who has not enrolled one is refused, not waved through.
• Administrative actions — including access granted under consent — are recorded.

Application hardening

• State-changing requests carry CSRF protection.
• Sensitive endpoints are rate limited, and endpoints that could otherwise confirm whether a private whiteboard exists return the same response whether it does or not.
• Files are stored in a private bucket and reached through short-lived signed URLs, not public paths.
• Payment details never reach our servers; checkout runs on Stripe.

Sub-processors

Five handle the content you write: Vercel hosts the app, Supabase stores it, Anthropic and OpenAI run the extraction that builds the Continuity Index, and Resend sends the weekly digest, which quotes the week’s decisions. Four more see far less: Stripe for billing, Sentry for server errors, GitHub for a bug report we choose to investigate (Anthropic’s model reads the report in that run), and Google Workspace for our own inboxes. The Privacy page sets out exactly which and why, with each company’s legal name and country. Your data is stored and the application runs in the United States; there is no choice of region today.

Encryption, backups and testing

Customer data is encrypted at rest with AES-256 and in transit with TLS, according to the published statements of our database and hosting providers — we rely on their controls rather than operating our own; the site sends HSTS and an enforced Content-Security-Policy. The database is backed up daily by our database provider and each backup is kept for seven days, so the most the database could lose is a day. Uploaded files are not in those backups: deleting a file removes the only copy, so keep your own. Our admin area requires a second factor on every session for our own staff. Secrets live in the hosting provider’s encrypted settings, never in source control, on a documented rotation schedule. Dependencies are scanned on every build, and a new high or critical advisory fails the build.

What we do not claim

We do not hold SOC 2, ISO 27001 or any other third-party attestation today, and we are not going to imply otherwise on a marketing page. We have not yet commissioned a penetration test. We do not yet offer a second factor for customer sign-in, single sign-on (SAML) or SCIM provisioning, a choice of where data is stored, or a way to switch off the AI processing for one organization. We do not offer a contractual uptime guarantee on self-serve plans, and a restore from backup has been documented but not yet drilled. If your procurement process needs any of these, talk to us before you start rather than after.

Reporting something

Found a vulnerability? Write to security@echo-logic.ai. Tell us what you found and how to reproduce it, give us reasonable time to fix it, and please do not access data that is not yours while proving the point. We will not pursue researchers who act in good faith.