VolShield Guard

Catches hidden instructions before your AI reads them, and pauses risky sends. Local-first prompt-injection shield.

VolShield Guard is a local safety layer between you and AI. It warns you when a webpage, file, or pasted text hides instructions that could mislead your AI, and it pauses risky sends before they happen. Nothing is sent to VolShield Guard. Your AI assistant reads everything on a page, including text you cannot see. Someone can hide instructions in white-on-white text, invisible Unicode characters, look-alike letters, and page comments to change what your AI does next. VolShield Guard catches those hidden instructions and explains them to you, in plain language, before your AI acts on them. Everything runs on your device. No account, no cloud, no AI API, no tracking. Free. WHAT IT DOES • Scans AI chat pages for hidden text, invisible characters, instruction-style wording, off-task commands, and data-exfiltration attempts. Five local checks, in milliseconds, and nothing is sent to us. • Sees the tricks that hide in plain sight: text tucked into hidden HTML attributes and metadata, and look-alike (homoglyph) letters from other alphabets used to disguise a command as ordinary words. • Scans PDFs for hidden instructions, on your device. Drop in a PDF, like a résumé or a contract, or scan one you're already viewing, and Guard checks its text for hidden, off-page, and tiny instructions. The file is never uploaded and its contents are never saved. The whole check runs on your device. • Shows a calm banner when something matters, and a side panel that explains every finding: what it is, why it was flagged, and where it is (a "Show on page" highlight for web pages, a page number for PDFs). Decoded hidden text is displayed as inert, quarantined text. • Guards the chat box itself. If you paste or forward text that carries hidden instructions, you get a warning right at the input, and risky sends ask for one explicit confirmation. You choose how often it pauses: on every risky send, only on the most serious ones, or never. You always stay in control: allow, override, or allowlist a site. • Offers a "Copy AI-safe version": a cleaned copy of the text, wrapped so your AI treats it as untrusted content instead of instructions. • Stays quiet on normal pages. Hidden text alone is normal on the web (accessibility labels, collapsed menus). VolShield Guard only warns when a hidden spot also contains something suspicious, so benign sites read "Looks clean." • Adjustable sensitivity, so you choose how cautious it should be. WHERE IT WORKS Nine popular AI chat services are protected out of the box, including ChatGPT and Claude; the full list is on volshield.com/guard and in the extension itself. You can also switch on protection for any other site you choose, one grant at a time. It is off by default, and nothing beyond the AI sites is accessed unless you turn it on. Detects instruction phrasing in five languages: English, Hebrew, Spanish, Russian, and Arabic, with full right-to-left support for Hebrew and Arabic. YOUR DATA STAYS YOURS • All detection happens locally in your browser. VolShield does not transmit or store your prompt content. Page text, chat-box text, and PDF contents are held in memory for the moment of the check, then dropped. • Nothing is sent to VolShield or any third party. There is no analytics and no telemetry. The one request that exists: if you ask it to check a PDF you already have open, it fetches that file from the address your browser already loaded, so it can read it. That goes to the site the file is on, not to us, and a PDF you drag in makes no request at all. • History is metadata only: the site hostname, the risk level, which checks matched, and the time. Never the text itself, never full URLs. Last 100 entries, clearable anytime. • No accounts, no sign-in, no remote code. The full privacy policy is short enough to actually read, and volshield.com/guard/security explains exactly what runs where and what is stored. WHAT IT CANNOT DO VolShield Guard reduces risk. It does not catch every attack, and no tool can; prompt injection is an unsolved problem across the industry. Think of it as a seatbelt for your AI: it catches the high-confidence tricks, shows you the evidence, and leaves the decision with you. For PDFs, it reads the text layer of a file. It cannot yet read text inside scanned images, and it cannot open password-protected PDFs. When a file has no readable text, it tells you plainly instead of calling it clean. On web pages, it does not yet read text inside embedded frames, content some sites build in a way that hides it from other extensions, or anything that is a picture rather than text. Those are on the roadmap, and we list them here rather than let you assume a page was fully checked. VolShield Guard is a VolShield product (volshield.com).