SquareX disclosed a new class of attacks called AI Sidebar Spoofing, where malicious browser extensions render pixel-perfect replicas of trusted AI sidebars (Comet and consumer browsers with AI features) and return crafted AI responses that trick users into executing harmful actions, from entering credentials on phishing pages to running commands that enable device takeover and data exfiltration. These extensions can stay dormant and behave normally until they detect an opportunity to deceive the user. SquareX’s research shows the attack abuses the UI/agent model of AI sidebars: a malicious extension injects or overlays a convincing fake sidebar (or hijacks sidebar responses) so users believe they’re interacting with a legitimate AI agent. Because modern AI sidebars are designed to complete multi-step tasks, users tend to follow procedural instructions; attackers substitute benign instructions with phishing links, clipboard/command prompts, or stepwise directions that cause credential disclosure, OAuth abuse, remote command execution, or installation of additional payloads. The attack succeeds with only common extension permissions and standard DOM/network capabilities (content scripts, DOM manipulation, redirecting links, background network calls), making permission scans insufficient to detect it. SquareX demonstrated cases including swapped exchange URLs (cryptocurrency credential theft) and malicious command sequences that enabled ransomware-style takeover. The research emphasizes that the vector affects standalone AI browsers and mainstream browsers that implement AI sidebars (Edge, Brave, Firefox, Safari).
Microsoft is currently addressing an issue that has been disrupting access to several Defender XDR portal functions for roughly ten hours. A service alert in the Microsoft 365 admi...
Ukrainian-aligned hacktivist groups have carried out a coordinated spear-phishing campaign against Russia’s aerospace and defense ecosystem, according to a recent Intrinsec threa...
Security researchers have uncovered a critical command-injection vulnerability in OpenAI’s Codex CLI, identified as CVE-2025-61260, which allows attackers to execute arbitrary co...