How your data is protected

Personal details are hidden before any AI sees them

Before a brief, a document or a table goes to any AI model, Yellowjacket swaps identifying details for placeholders: names of people and organisations, email addresses, phone numbers, street addresses, bank, card, tax and ID numbers, web addresses, IP addresses and secret keys. "Acme Pty Ltd" becomes {ORG_1} everywhere it appears, so the models can still match it across documents. The list that maps placeholders back to the real details stays in memory on our server for that one job, and the real details are put back in your answer before you see it. Amounts, dates and invoice or order numbers are kept, because the checks need them.

This is on for every workspace unless you switch it off (Settings, Privacy). More on how it works.

Where it doesn't apply:

• Photos and scanned pages are sent as pictures; nothing reads the pixels to hide names first.
• Requests for a picture, questions about your connected data, apps and software builds are sent as written.
• In questions and AI debates, company and product names stay visible so the AI knows what you mean. People's names, contact details and numbers are still hidden. Well-known public names (products, platforms, large public companies) are never hidden anywhere.
• To answer a question about something it doesn't know, the AI may search the web for public names. Only those search phrases go to the search, not your question.

No detector finds every name. Hiding details lowers what the AI providers see; it does not make it zero.

AI providers that keep nothing

Every request to an AI model goes through OpenRouter and asks it to use only providers that keep no copy of the request and don't use it for training (OpenRouter's zero data retention setting, with data collection denied). That is the providers' commitment under OpenRouter's policies; we can't inspect their systems ourselves. Models that have no such provider are left out.

We don't use your documents or results to train AI models. That is off for every workspace and would only be switched on with your agreement.

Secrets and passwords

• Encrypted at rest: the AI provider key we create for each workspace, the sign-in tokens of apps you connect, and the addresses and passwords of databases you connect are stored encrypted (Fernet, with a key held by the server).
• Sign-in codes are stored only as a keyed hash, last 10 minutes, allow 5 tries and work once.
• Passkeys: we keep only the public key. The private key stays on your device or password manager.
• Card details go to Stripe and never reach us.

Your databases and links

• Databases are read, never written. Where the database supports it, the connection itself is set to read-only. On top of that, each query must be a single read-only SELECT, has a time limit and returns at most 5,000 rows. We still recommend connecting with a read-only database user.
• No reaching into private networks. Connections to private or internal network addresses are refused, so a connection string can't be used to probe our network.
• Links you add are fetched only from public addresses, on the normal web ports, up to 15 MB, with a 20-second limit.
• Read-only connectors for apps such as Google Drive, SharePoint and Xero are coming soon; they are not available yet.

Signing in

You sign in with a one-time code sent to your email address, or with a passkey. Codes and passkeys mean there is no Yellowjacket password to steal or reuse. Once you're signed in, your browser holds a random session token in a cookie that page scripts can't read (HttpOnly), sent only over HTTPS; we store only a hash of it. A session lasts 90 days and renews while you keep using it.

Sending a sign-in email and opening a new workspace are protected by Cloudflare Turnstile, a bot check that is invisible for almost everyone.

Apps and code the AI writes

• Apps are previewed in a sealed, sandboxed origin: they can't read your Yellowjacket session, can't send requests to other sites and their forms can't post anywhere. Their automatic checks run in a headless browser with the network blocked.
• Software builds: the tests the AI writes run in a separate process with memory and process limits and, where the server allows it, with no network, as an unprivileged user that can't see the server's other processes or its data folders. Each build records which of these protections it ran under.

This is containment, not a separate machine: the code runs on the same server (and the same operating system kernel) as the rest of the service. Review generated code before you run it yourself.

Where it runs and backups

Yellowjacket runs on Cloudflare, and every connection is over HTTPS. The service's data (accounts, work orders and results, saved plans, encrypted connection details) is backed up to Cloudflare R2 storage every few minutes when something has changed, with a dated copy kept each day. Original files you upload are not included in backups, except while a job is waiting for credit to start.

What we don't claim

We don't hold any security certification or audit (such as SOC 2 or ISO 27001), and Yellowjacket is not set up for health records under rules such as HIPAA. Everything on this page describes how the code works today, and we say where it falls short.

Found a problem? Email hello@aiyellowjacket.com. How we handle personal information is on the privacy page.

Related