Microsoft Flags AI-Driven Phishing: LLM-Crafted SVG Files Outsmart Email Security

According to Microsoft Threat Intelligence team, a new phishing campaign aimed at U.S.-based organisations has likely used large language models to obfuscate payloads inside Scalable Vector Graphics files and evade security tools. The activity, detected on 28 August 2025, employed an already compromised business email account to send lure messages that masquerade as a file-sharing notification, urging recipients to open what appears to be a PDF but is in fact an SVG file.

The attackers used a self-addressed email tactic, with recipient addresses matching and targets hidden in the BCC field to bypass basic detection heuristics. Microsoft notes that SVGs are attractive to attackers because they are text-based and scriptable, enabling JavaScript and other dynamic content to be embedded directly within the file, allowing interactive phishing payloads that can bypass some security checks.

The campaign’s core function redirects victims to a landing page, triggers browser fingerprinting, and initiates session tracking, with obfuscation aided by business terminology and a lengthy narrative inside the SVG code. Microsoft said the overall campaign was limited in scope and effectively blocked, though such techniques may be adopted by other threat actors.