Kerb

Checks prompts in your browser before they reach ChatGPT, Gemini or Claude, and warns you about secrets or personal data.

Kerb is an AI-DLP (Data Loss Prevention) extension that inspects text you're about to send to ChatGPT, Gemini, and Claude entirely within your browser. If sensitive information is detected, a warning dialog appears before the message is sent. ■ Detection Categories (37 rules active in this release) · Security credentials — 18 rules (cloud API keys, access tokens, private keys, DB connection strings, etc.) · Personal identifiers — 12 rules (My Number, credit card numbers, passport numbers, national ID numbers, IBAN, bank account numbers, etc.) · Confidential content — 7 rules (confidentiality markers, export-control/sanctions terms, salary figures, M&A information, source code, etc.) Digit-count matches alone do not trigger a warning: credit cards are checked with Luhn, IBANs with mod-97, and My Number with its check digit. ■ Privacy-First Design · All text analysis runs locally in your browser — no cloud processing · Raw data is never sent to Kerb's servers — stored only in the storage you connect (currently Dropbox) · While no storage is connected, logs are encrypted with AES-256-GCM in your browser and flushed once connected · Choosing "Edit text" stores nothing, because the message was never sent · Kerb's operators never store or view your input ■ Key Features · Warning dialog with "Edit text", "Report and send", and "Send anyway" · Kerb never blocks a send. It warns you; the decision is always yours · Attachment text inspection (text files and PDFs) · If an attachment could not be read, Kerb tells you before sending even when no rule matched · Popup statistics: total warnings, category breakdown, action breakdown · Kill switch via signed CDN config (disable specific rules remotely) · Japanese / English UI ■ About Attachments (important) Files you attach are uploaded to the AI service the moment you attach them, before you press send. That is how those services work, and no extension can prevent it. What Kerb can stop is sending a message that uses the file. When a warning involves an attachment, the dialog says so. ■ Supported Sites · ChatGPT (chat.openai.com / chatgpt.com) · Google Gemini (gemini.google.com) · Claude (claude.ai) ■ Log Destination · Dropbox (OneDrive / Google Drive / Box coming later) · Detection and warnings work with no storage connected ■ Required Permissions · storage: Save local encrypted logs and settings · unlimitedStorage: Retain pending logs beyond the default ~10MB cap · alarms: Schedule hourly config refresh · identity: Dropbox integration (OAuth) · chatgpt.com / chat.openai.com / gemini.google.com / claude.ai: Read the text you are about to send and show the warning dialog · config.kerb.zone: Fetch the signed rule configuration (read-only; your input is never sent) · api.dropboxapi.com / content.dropboxapi.com: Dropbox authentication and log upload ■ On False Negatives Kerb matches known patterns and does not guarantee detection of every kind of sensitive information. The final judgement and responsibility remain with you.