Skip to main content
These are the questions security teams ask us most. The answers are short and checked against our code. Where something isn’t done yet, we say so and give the issue number we’re tracking it under.

Encryption

No. Messages are scrambled on your device and unscrambled on your agent’s computer. Our servers store and pass along the scrambled version, and the keys aren’t on our servers.Two honest limits. First, your agent reads your messages and sends them to its own AI provider, which sees them. Second, a server that’s been taken over could add a member to a room or chat today, and that member could read new messages. We’re fixing that (issue #1321). See the overview.
Two systems, both well known.MLS (Messaging Layer Security) is our main one. It’s an internet standard (RFC 9420) built for group chats, and we use it for every room and for 1:1 chats. It scrambles messages with AES-128-GCM, a widely used cipher. Keys are agreed with X25519, and messages are signed with Ed25519.Double Ratchet is a backup for 1:1 chats only. It covers the gap when a new chat’s MLS setup hasn’t arrived yet. It scrambles messages with XChaCha20-Poly1305, a modern cipher with a long random number per message that makes accidental reuse very unlikely.Files and history backups also use XChaCha20-Poly1305. Full detail is in Technical details.
Partly, and we won’t oversell it. MLS changes keys as members join, and each message gets its own key. But your devices keep a short window of older keys so late messages can still be opened. Your agent also keeps a readable copy of recent messages on its computer, and copies of your own sent messages use one fixed key (issue #1324).MLS can also heal after a device is compromised, but only when members refresh their keys. When you remove a device, your web app now changes the room’s keys to cut it out. But members don’t refresh their keys on a schedule yet. So we don’t claim post-compromise security today.
Yes. A notification never includes what anyone wrote. A new message says “New message from <agent>”. A decision says “<agent> needs a decision”, never its title or choices.Room messages never send a push. A 1:1 chat usually sends one only if a message waits a while without your app picking it up, up to about 15 minutes. A message sent the older 1:1 way can send one right away.One alert does carry a name you chose: when a device is waiting to join a room, we tell you “Someone is waiting to join <room name>.” So Apple, Google and our push provider can see that room’s name.
Yes. The app talks to our servers over HTTPS through Cloudflare. We haven’t yet verified that the last hop, from Cloudflare to our server, is encrypted too (issue #1325). Message content is scrambled end-to-end either way.

Keys and devices

  • Phone app: in your phone’s secure storage.
  • Web browser: in the browser’s local storage, not encrypted. Anyone with access to your browser profile can read them.
  • Your agent’s computer: in a file only the agent’s user account can open.
For your most sensitive conversations, use the phone app. See where your keys live.
You build the agent in the app, which gives you a one-time command. The command holds a connection code that works once and expires in 24 hours. We store only a scrambled fingerprint of that code, called a hash, not the code itself.The agent makes its own keys on its own computer and sends us only the public half. The app then shows you a short code, called a fingerprint, made from the agent’s public key. If the agent’s terminal prints one too, check that they match, then tap Approve & Chat. Nothing reaches the agent until you approve.
Today, a new phone or browser is trusted as soon as you sign in. No existing device has to approve it. That means anyone who takes over your sign-in gets a trusted device. We’re adding device approval (issue #1322).Until then, you’ll hear about every new sign-in. Your other devices get a New device signed in alert, and we email you. Both lead to Linked Devices. There you can tap Sign out this browser, which keeps everything it wrote, or Remove and erase, which also deletes the messages it sent. Either way it’s signed out right away, and your chats and rooms keep working on your other devices. If you tap This wasn’t me — remove and erase, the rooms that browser started are deleted too.Either action also ends that browser’s sign-in, so a window that’s still open goes to the sign-in screen.Three things can leave a sign-in open. We can’t end a sign-in we never recorded, so a browser that hasn’t connected since we added this can come back as a new one. We leave a sign-in alone if a browser you’re still using shares it, which can happen after you clear a browser’s data. This wasn’t me ends it anyway, unless it’s the sign-in you’re using right now. And if our sign-in service can’t be reached at that moment, the sign-in stays open there until we get through, though AgentVault refuses it right away. We keep trying for seven days. Changing your password ends every other sign-in, so it closes all three.Then change your password. Whoever got in may still know it. Do the same if one of your own browsers is signed out and you didn’t do it, because anyone signed in to your account can sign out your other browsers.You can see your devices under Settings › Devices › Linked Devices.
Yes. Settings › Security › Key Backup scrambles your keys with a backup code and uploads them. The app shows the code once. Without it, we can’t open the backup, and neither can you.Your private keys never leave your device unencrypted.

Removing an agent or device

You have a few options, from pause to delete. See Fix an agent.When you Revoke an agent, our servers act right away. They:
  • mark it revoked, so it can’t sign in again,
  • close its chats,
  • take it off every room’s member list, so it gets no new messages, and
  • cut its live connection.
Yes, once your web app catches up. Removing an agent stops our servers from sending it anything, right away. Then your web app changes each room’s keys so the agent is out. That happens at once for rooms you have open, and otherwise the next time one of your browsers opens the room or chat.Until that happens, the agent could still open new messages it got some other way. That would take a compromised server or a copy of our traffic. Only the web app does this today, not the phone app or your agents (issue #470).
No. Deleting in the app removes the agent and its data from AgentVault. Its keys and readable message history stay on its computer until you remove them there. Fix an agent has the clean-up commands.

Tenant isolation

Two ways. Every query in our server code is limited to your account. And since July 24, 2026, the database itself enforces the same rule, using a PostgreSQL feature called Row-Level Security. It filters every row by account, even if a query forgets to.A few things to know:
  • The app connects to the database with a role that can’t skip those rules.
  • A separate, more powerful role is used for some system work and our admin console. It can see across accounts.
  • A few internal tables have no account column, so Row-Level Security doesn’t cover them.
  • Row-Level Security separates accounts. It doesn’t decide who’s in which room. That’s checked in our server code.
Our admin console can see across all accounts. Staff can see things like account details, agent names and room member lists. They can’t read your message content. See What AgentVault can and can’t see.

Data location, retention and deletion

In the United States. Our server, database and file storage run at DigitalOcean in US data centers. Attachments are stored scrambled in DigitalOcean’s New York region.The web app’s files are served by Vercel, and all web traffic passes through Cloudflare. The full list of services is on What we can see.
We keep your scrambled messages and files until you delete them. There’s no automatic time limit today.
  • Delete an agent to remove it and its chats.
  • Delete your account under Settings › Delete Account to remove everything in it: agents, chats, keys and settings. Only the account owner can do this, and only from a signed-in app, not with an API key or an agent.
There are two codes worth keeping safe.
  • Your Key Backup code unlocks your backed-up keys.
  • Your history recovery code unlocks your agent’s message history backup, so its past messages can follow it to a new computer. See History recovery.
Anyone with a code can open what it protects. We can’t reset or recover these codes for you.

Sign-in and access

In a limited way. When sign-in asks for a second step, it’s a code sent to your email. We don’t support authenticator apps, passkeys or security keys yet. See Account and settings.
Everything you can do. An API key acts with the full permissions of the person who made it. There are no limited keys yet.So treat it like a password. Store it somewhere safe, don’t paste it into chats, and revoke it under Settings › Developer › API Keys if it might have leaked. We store only a scrambled fingerprint of each key, and the app shows the full key once, when you create it. More in the developer docs.
Yes, for the actions most open to abuse:One honest note: these limits rely on our cache server, and they switch off if it’s down.

Reporting a security problem

Contact us through the app. Tap Help (it’s also under Settings › Support), describe the problem in a sentence, then tap Send to support. A person will reply by email.The Help assistant isn’t end-to-end encrypted, and its AI provider sees what you type. So keep that first message short. Don’t include exploit details or secrets. We’ll reply with a better way to share them.We don’t run a public bug bounty yet.

What’s next