AI-Generated SVGs: The Phishing Attack Designed to Bore Your Defenses to Death
Attackers are using LLMs to write obfuscated phishing code hidden inside SVG files, disguised as boring business reports to evade security filters.
This post was originally published on LinkedIn.
TL;DR
- The Scam: Attackers are using AI, specifically Large Language Models (LLMs), to write the code for their phishing attacks.
- The Method: Hiding this malicious code inside Scalable Vector Graphics (SVG) files - image files that can secretly run scripts.
- The Trick: Using AI to bury the malicious parts in a sea of corporate jargon, making the code look like a harmless business report.
- The Goal: Same as it ever was - to get you to a fake login page and steal your credentials.
Your email filter is smart. AI-driven phishing is smarter.
Let’s be honest - for years, the saving grace of email security has been that most hackers write phishing emails with the grace of a drunken ex sending DMs at 4 AM. Bad grammar and weird phrasing required very little attention to spot as a threat. With ubiquitous, easy to access AI, that era is over.
According to a recent dispatch from Microsoft’s Threat Intelligence team, we’re now dealing with phishing campaigns where the malicious code is likely written by an AI - which is not surprising - the clever bit is that the attackers have shifted their focus from crafting a more believable email to engineering a perfectly invisible payload instead.
The new trick - hiding a missile in a quarterly report
The attack Microsoft broke down is a masterclass in deception. It starts with a compromised email account and targets users with a self-addressed message, hiding the real victims in the BCC line to bypass basic security checks. The email urges you to open what looks like a PDF, but it’s actually an SVG file.
Why an SVG? Because it’s the perfect Trojan horse. It’s an image format that browsers render, so it seems harmless. But it’s also text-based and can contain JavaScript. Something like the old ‘macros in Excel file’ but now hiding in a vector image.
But here’s where the AI-generated genius comes in. When security analysts (or automated filters) inspect the file’s code, they don’t see obvious malicious scripts. Instead, they see what looks like code for a business analytics dashboard. The AI has obfuscated the payload’s real function - to redirect you to a credential-harvesting site - using an absurdly long and complex sequence of business-related terms. Function names and variables include words like revenue, operations, risk, quarterly, and growth. All incredibly tedious.
It’s like trying to find a single poison pill in a 500-page corporate filing. It’s designed to be so boring and so legitimate-looking that no one bothers to check page 497.
The ghostwriter in the machine
The code was so bizarrely over-engineered that Microsoft’s own AI, the Security Copilot, flagged it as something a human probably didn’t write. The verbosity, the redundant naming, the formulaic structure - it all pointed to an LLM being fed a prompt like: “Write me some SVG code that redirects a user, but make it look like a boring business report to avoid detection.”
This is a massive shift. The barrier to entry for creating highly evasive, sophisticated malware just dropped to the floor. Any threat actor with access to a decent LLM (in other words, an internet connection and a web browser) can now generate complex, obfuscated code that can slip past traditional defenses.
The obvious tells that made phishing attempts laughable are gone. You can’t just “look for bad grammar” when the lure is written as well as your own company’s marketing copy and the payload is disguised as a P&L statement. In this new arms race, the most dangerous file you’ll receive is the one that looks completely, professionally, and perfectly normal.