Remora

Remora - AI Agent Prompt Injection Detector Remora silently scans every page you visit for hidden prompt injection attacks, malicious instructions embedded in web pages that attempt to hijack AI agents browsing on your behalf. How it works When a page loads, Remora scans the DOM for 8 types of hidden instruction techniques: invisible text, off-screen elements, color-matched text, tiny fonts, aria-hidden abuse, HTML comments, suspicious meta tags, and CSS content injection. If injections are found, a blocking overlay appears with a 30-second countdown. A human can click "I acknowledge this risk" to dismiss it. An AI agent cannot — when the timer expires, a desktop alert fires and the page stays blocked. What it detects • Instruction overrides ("ignore previous instructions", "new instructions") • Role manipulation ("you are now", "act as") • Data exfiltration commands ("send conversation to", "reveal system prompt") • Action hijacking ("instead output", "do not complete") • Conditional triggers ("when the user says", "codeword") New in v0.2 • Pineapple check - a visible human/AI discriminator test in the overlay • Instant desktop notification on detection (click to jump to the affected tab) • Suspicious domain blocking — catches brand impersonation and known phishing sites • Attack category explanations showing why each detection is dangerous • False positive reporting link built into the overlay New in v0.2.1 • "Don't scan this site again" option on all warning overlays • Trusted Sites panel in the popup — view and remove trusted sites at any time • Fixed false positives on major brand domains • Improved warning overlay footer design New in v0.2.2 • Fixed an issue where a single hidden instruction could be reported multiple times on certain pages • Fixed a minor visual alignment issue in the warning overlay's details view • General security hardening Zero data collection. All scanning happens locally in your browser. Nothing leaves your device.