AgentLens — See What Agents See

See machine-readable webpage and PDF content that is not normally visible to you.

AGENTLENS — SEE WHAT AI AGENTS CAN READ AgentLens is a privacy-first browser extension that detects prompt injection, hidden instructions, and machine-readable content on web pages and supported PDFs. AI assistants and browser agents may encounter content that is difficult—or impossible—for you to see. This can include transparent text, clipped or off-screen elements, hidden containers, accessibility attributes, metadata, HTML comments, PDF annotations, and invisible PDF text layers. Attackers can misuse these surfaces to insert instructions that attempt to override an agent’s task, request credentials, invoke tools, expose sensitive information, or send data somewhere unexpected. AgentLens helps you inspect that content before copying it into an AI chat, asking an agent to summarize it, or allowing an automated tool to use it. WHAT AGENTLENS CAN DO • DETECT PROMPT-INJECTION ATTEMPTS Scan the current page for signals such as: – Instruction overrides and role impersonation – Credential or secret requests – Data-exfiltration instructions – Commands to invoke tools or contact external services – Prompt delimiters and injected system-style messages – Concealment techniques and obfuscated instructions • REVEAL HIDDEN AND MACHINE-READABLE CONTENT AgentLens identifies content that an AI extraction tool might encounter even when it is not normally visible to you. It highlights the nearest visible source location and presents concealed text in a readable overlay. Previous and next controls let you move through every detected finding. • EXPLAIN EVERY FINDING AgentLens does not rely on an unexplained risk score. Each finding can show: – What was detected – Why it was flagged – How the content was concealed – Which extraction profiles could encounter it – Where it appeared on the page – Which deterministic rule triggered the finding – What action is recommended • DISTINGUISH SUSPICIOUS INSTRUCTIONS FROM ORDINARY HIDDEN CONTENT Not all machine-readable or hidden content is malicious. Responsive menus, accessibility labels, metadata, and other routine page elements can be legitimate. AgentLens separates suspicious agent-directed instructions from neutral visibility differences so you can evaluate the evidence yourself. • REMOVE SUSPICIOUS SOURCES FROM THE CURRENT PAGE Strip individual detected sources (or all detected sources) from the working page before copying or passing its content to an AI tool. These changes are temporary. Reload the page to restore its original content. • INSPECT A COMPLETE SCAN REPORT Open the full report to: – Review findings by risk level – Inspect the detected evidence – Copy source locators – Reveal findings on the original page – Strip individual sources – Export a bounded JSON report • ENABLE LIVE PROTECTION Optionally scan pages after navigation and receive an alert when medium- or high-risk instruction signals appear. You can allow live protection on all websites or restrict automatic access to websites you choose. • SCAN SUPPORTED PDFs AgentLens can inspect digital and OCR-generated PDF text layers for: – Invisible text – Clipped or off-page content – Suspicious annotations – Metadata – Agent-directed instructions AgentLens does not perform OCR on image-only PDFs. • CUSTOMIZE THE DETECTION RULES Review the active, versioned ruleset, edit detection behavior, and import or export rules for reuse, testing, and auditing. • KEEP AN OPTIONAL LOCAL HISTORY Revisit previously scanned pages without creating an account or uploading your browsing history. Local history can be disabled or cleared at any time. PRIVATE BY DESIGN Scanning happens locally in your browser. – No cloud scanning – No AI model required – No telemetry – No account – No remote rule downloads – No permanent scan history unless you enable it – No report export unless you explicitly request it AgentLens requests website access only for the protection mode you select. Manual scans use the access granted through your toolbar action. WHEN TO USE AGENTLENS AgentLens is designed for people who: – Copy webpage content into AI chats – Ask AI assistants to summarize or research websites – Use browser agents, automation, or tool-enabled assistants – Review unfamiliar websites and documents – Test prompt-injection defenses – Audit what a webpage exposes to machines – Need inspectable evidence instead of a black-box warning TRY AGENTLENS SAFELY These test pages contain examples of hidden content and prompt-injection techniques: – https://ref.jock.pl/modern-web/ – https://garden-advice.netlify.app/ – https://www.agentpwn.com/attacks IMPORTANT LIMITATIONS AgentLens does not claim to reproduce the exact behavior of every AI assistant or determine that a page is completely safe. It compares multiple machine-readable extraction profiles and applies transparent, deterministic rules. AgentLens gives you evidence and remediation controls; you make the final decision about what an AI tool should be allowed to read. Install AgentLens to reveal hidden instructions before they reach your AI workflow.