Researchers demonstrated a prompt-injection method that hides malicious instructions inside encrypted text, then tricks an AI assistant into decrypting it using built-in code tools. In tests, a normal “summarize this page” request could cause Grok to exfiltrate chat data without any click or warning, and could push Gemini into producing content it would normally refuse.
Mirage safely runs attacks like this one against your own team, so you find out what happens before a real adversary does.
Imagine you just type: “summarize this page”... and that alone makes the AI leak your past chats. Researchers call this “Cryptographic Context Injection”, malicious instructions are hidden in encrypted text on a webpage, then the AI quietly uses its code tools to decrypt and obey them. In tests, Grok turned that simple request into silent chat-data exfiltration, and Gemini into content it normally refuses, because by the time the text is decrypted, it’s already past the guardrails. Your move: when an AI is summarizing an unfamiliar link, don’t type in secrets, and if it suddenly wants to decrypt, run code, or open extra links, stop and close that chat.
Researchers found a real phishing campaign offering a “free” Claude Max upgrade to trick people into signing in with Google. The site uses a fake, draggable Google login pop-up (“browser-in-the-browser”) that looks legitimate and captures credentials. A stolen Google account can expose email and…
Researchers found a phishing campaign offering a “free” upgrade to Claude Max to trick people into signing in with Google. The page uses a convincing fake, draggable Google login window (“browser-in-the-browser”) to capture credentials, potentially giving criminals access to email, documents, and…
Researchers demonstrated a technique where encrypted instructions bypass AI safety filters because the filters can’t interpret the ciphertext. In examples against xAI’s Grok and Google’s Gemini, the model is tricked into decrypting attacker-supplied text inside a trusted environment, enabling data…
This bulletin describes multiple real-world scams where attackers make fake pages and messages look like routine, trusted experiences (search answers, Google login pop-ups, “giveaways,” and official-sounding calls). Examples include a fake Claude Max giveaway using a convincing fake Google sign-in…
Researchers described an Android Trojan (“RatHat”) that starts with scam texts or malicious ads and tricks people into installing a fake app from a bogus download page. After installation, it pressures victims to grant Accessibility permissions using fake excuses or incentives, then uses those…
Malwarebytes reports that a multi-stage Windows malware loader called PavinLoader is being delivered through multiple real-world campaigns, including ClickFix “verification” pages and fake software downloads. Victims are tricked into running installers or scripts that use legitimate Windows tools…
