Clotr: AI chat privacy, clot your data leaks

Warns you before a password, a card number or a personal detail goes into an AI chat.

Clotr checks your message on this device before you send a password, a key or a personal detail to an AI chat. Clotr has no server. If you choose to send the message, it still goes to that AI service as usual. What it looks for - Passwords, API keys, tokens, connection strings and wallet seed phrases - Phone numbers, emails, addresses in the US, UK, Canada and Australia, and dates of birth, however they're typed, misspellings included - Card numbers, card security codes, gift card numbers, SSNs, bank accounts, IBANs, UK, Canadian and Australian account numbers, and several national ID numbers - Sign-in and backup codes that scammers ask you to read out, with a line saying who would really ask for one - Your own list, like your name, your family and your street, which Clotr keeps only as a scrambled code - Text inside files you attach, photos that still carry the place they were taken, and photos or files named for a seed phrase or backup codes. Clotr takes the place out before the site gets the photo. It's on for 19 AI chats out of the box. Email and chat sites, search boxes and support chats are off until you tick them, on the welcome page or in Settings, and the browser asks you first. One more switch, "Everywhere I type", reaches a chat-style box on any website at all, beyond the ones Clotr already knows. It's off until you turn it on, and even then Clotr only looks at a box that reads like a chat. A form field holding your name, address, card or password is left alone. If an AI reply mentions a detail from your vault, Clotr shows a count on its toolbar badge and in the popup. Activity keeps the kind of detail and chat name; the page sees no notice, and Clotr never keeps the reply or detail. When Clotr finds something, a warning appears in the corner of the page and names it. Choose Clot this leak to cover the detail, or Leave it alone to close the warning and leave the message in the box; then send it with the site's usual Send button. For a kind you've set to Ask before sending, Leave it and send sends the held message after you choose it. Scam check looks out for you too. If a command you copy on an AI chat looks like the "verify you're human, paste this" trick, Clotr warns you before you run it and offers to clear it from your clipboard. Paste a message that worries you into "Is this a scam?" and it marks the signs it knows, in English or Spanish, and explains what each one means. It says plainly that a message it finds nothing in can still be a scam. Tourniquet gives stronger settings to someone you set Clotr up for. After a scam or a near miss it can ask before money or an ID number goes out for 30 days, then step back down on its own. Practice lets you try spotting both safely, with made-up scam messages checked by the same engine that protects you, and made-up AI chat replies that show what would leak. A link you paste can carry a code you didn't mean to share, like one an app adds when you copy a share link, or a tag an ad network uses to follow the click. Clotr warns you by name and offers to strip just that part, with a page that explains what the code does and what Clotr can and can't see about it. Privacy Clotr doesn't use AI to check your message. It uses plain pattern matching in your browser, and keeps a short history in this browser profile on your device. Clotr has no server for chat text or findings; if you choose to send a message, the AI service gets it as usual. You can export or delete the history. The code is open source under the AGPL-3.0-or-later, and the build is reproducible, so you can check that the store's zip matches the source. Chrome lists what Clotr reads on your computer. None of it is ever sent anywhere. Why it asks for each permission - Storage holds your settings, your list as a scrambled code and the history, on this computer. The history keeps the kind, site, time and what you chose, and leaves the text out. - The AI chat sites let Clotr check what you type in their message box, show the warning there, and swap a detail out when you choose to hide it. - Active tab lets Clotr check one tab once, when you open it on a site that isn't a known AI chat, to tell you whether that site looks like one. It only happens on that click. - Scripting runs that one check. It also starts Clotr on a site you've added, or in AI chat tabs that were already open when it was installed or updated. - Declarative content puts a small dot on the toolbar icon when a page looks like an AI chat, without reading the page. - Alarms reset the toolbar's count for today after midnight, and remove history older than you chose to keep. - Other sites are optional. Clotr asks only for the email and chat sites, search engines and support chats you tick from a fixed list, or for every website once you turn on Everywhere I type, and the browser asks you first either way. Made by Alex, a computer science student. The code, the issues and how to build it yourself are at github.com/clotr-app/clotr.