Security
How Klyroe keeps you safe
Last updated 10 November 2026
Klyroe's agents browse, fill in forms and run tasks for you. This page lists the controls built into Klyroe to keep that in check, in plain terms. Items marked Coming soon are built but not switched on yet. If something here doesn't match what you see, tell us and we'll fix the product or this page.
1. Your own computer, walled off
When Klyroe needs a computer (to browse the web, run code or keep files for a task) it uses a small virtual machine that belongs to your account alone. It isn't shared with other people.
- Each machine runs on its own private network, so it can't reach Klyroe's database or our other servers.
- Your workspace is kept on an encrypted disk attached to your machine.
- The machine stops when it's idle and starts again when you next need it.
2. Your passwords stay with you
- The agent never types into password, one-time code or card number fields, and never types a card number anywhere.
- When a site needs you to sign in, the agent pauses and asks you to take over. You sign in yourself; the agent can't see what you type, and it waits until you hand back.
- Keys and tokens for the apps you connect are stored encrypted. Our servers use them to make the call; they aren't put into what the AI reads.
3. You approve what matters
The AI proposes; Klyroe's servers decide. When an action needs your OK, you see a card with exactly what will happen.
- Anything that submits a form, books, sends, posts, buys, accepts or deletes asks you first, one action at a time.
- Deleting things always asks, even if you've said “Always allow” for that kind of action before.
- Some pages are off limits to the agent whatever your settings: online banking and money transfers, payment and checkout pages, crypto, government ID services, and password and security settings. Payments are yours to make.
- You can approve an action once, for one chat, or for one site, and keep a short list of rules in Settings → Permissions. Nothing that moves money, sends a message, places a call or deletes data can be approved in advance, and once outside content has reached a task, approvals you gave earlier don't skip the card.
4. A check on what leaves
Before the agent opens a new site, fetches a page, or types into or submits one, Klyroe checks what would be sent.
- Addresses known for collecting stolen data (request catchers, tunnels, paste sites) are blocked.
- A password, key, sign-in token or card number in an address or in text to be typed is blocked: you type those yourself.
- Personal details you didn't give in the chat, or that a page asked for, need your OK before they go out.
5. A record of what Klyroe did
The Activity page lists every action Klyroe took or was refused for you, in chats, helpers, agents and background tasks. You can search it and export it. Passwords, keys and card numbers are taken out before anything is stored.
6. Web pages and files can't give orders
A web page, email or file can contain text written to trick an AI (“ignore your instructions and…”). Klyroe treats everything that didn't come from you as information, not instructions.
- Results from pages, files and tools reach the AI labelled as outside content.
- When that content contains instruction-like text, Klyroe tells you where it was found.
- Once outside content has reached a task, even small changes the agent would otherwise make on its own ask you first. Only Klyroe's servers can mark a task this way, and nothing the AI writes can clear it.
7. Memory you can see and control
In Settings → Memory you can see everything Klyroe remembers about you, edit or delete any of it, pin what matters, and export it all as a file (JSON or Markdown). You can turn learning from your chats off at any time. See the privacy notice for how memory works.
8. Your data is kept to your account
Every database table that holds account data has row-level security switched on and enforced: a request made as you can only reach your own rows, and a test checks every such table on each change. Your chats, memory and files are only used to improve Klyroe if you choose that (Improving Senova).
What Klyroe still sends out: your prompts and files go to the model provider handling the request, for that request only (listed on our subprocessors page). Organizations can also turn on masking of contact details and card and ID numbers before anything leaves for a provider; it is on by default for the developer API.
9. Learning from mistakes, with your permission
If you chose to help improve Klyroe, a task that fails (a step that broke, a thumbs down, or a task you stopped) is kept as a test case with your own name, contact details, card and ID numbers and secrets masked and links cut back to the website name. Only Klyroe staff can read these cases, and deleting the chat deletes its case. We publish how often tasks succeed, as counts only, on our task scoreboard.
10. Coming soon
- Phone calls that always say they're AI. When Klyroe phones a business for you, it opens by saying it's an AI assistant calling on your behalf, answers plainly if asked whether it's human, and ends the call if the person doesn't want to talk to an AI.
Questions? Email hello