Photo by Laine Cooper on Unsplash
Despite the headlines shouting about AI making cyberattacks easier and more dangerous, Google just dropped a reality check. They analyzed five recent malware samples created with generative AI, and the verdict? These so-called AI-powered threats aren’t doing much of anything.
Let’s break down what really happened—and why all the hype might be a bit premature.
Meet the “Failures”: AI Malware That Doesn’t Quite Work
Google’s security team looked into five malware families developed with help from AI:
- PromptLock
- FruitShell
- PromptFlux
- PromptSteal
- QuietVault
You’d expect scary, advanced cyberweapons, right? Not so much. These AI-built samples were sloppy, outdated, and easy to catch. Even the most basic antivirus tools could spot them. Most didn’t bother with advanced tactics like persistence or evasion. And none had any real-world impact.
One of the better-known examples, PromptLock, was actually created for an academic study. The study aimed to test whether large language models (LLMs) could handle planning and executing a full ransomware attack. But the result? According to researchers, it skipped major components like persistence or lateral movement—basically, it couldn’t hold a foothold or spread across systems. They ended up calling it a proof of concept, not a production-grade threat.
“You’d Ask for a Refund”: Experts Weigh In
Photo by Glen Carrie on Unsplash
Cybersecurity pros aren’t impressed. Independent researcher Kevin Beaumont put it bluntly: “If you were paying malware developers for this, you’d be furiously asking for a refund.” Another expert (who preferred to stay anonymous) shared a similar thought: Generative AI isn’t creating scarier malware—it’s just helping people repackage existing techniques.
That’s the core message Google’s report drives home. Even after years of AI hype, malware development using LLMs is crawling forward, not leaping ahead.
AI’s Real Role? Helping, Not Inventing
Yes, AI is being used by malicious actors. Companies like Anthropic and OpenAI have reported threat actors using their models—like Claude and ChatGPT—to assist with tasks like writing encryption code or debugging malware. But here’s the kicker: they’re not inventing new malware ideas. Just getting help troubleshooting or streamlining old ones.
Sometimes, these reports get breathlessly quoted without mentioning the fine print. In most cases, even the researchers themselves admit the limitations. No breakthroughs. No magic. Just AI making the same old tricks a little faster to pull off.
One Notable Loophole… and a Fix
Google’s report did highlight one clever trick. Someone managed to bypass guardrails on Gemini, Google’s own AI model, by pretending to be a white-hat hacker participating in a “capture the flag” cyber game. That kind of game is common in computer security circles, meant to test attack and defense tactics.
Still, Google caught on—and quickly adjusted its guardrails to close that loophole.
So Where Does That Leave Us?
Right now, AI-generated malware is more of a curiosity than a threat. To be clear, that doesn’t mean we can ignore it forever. AI will keep improving, and so will its ability to assist with attacks. But we’re far from a point where it can create malware better than human developers.
For now, the biggest threats out there? They’re still using old-school methods, not large language models.
Bottom Line
Here’s the takeaway: Generative AI isn’t building next-level malware—at least not yet. Google’s deep dive into five AI-generated malware families shows just how underwhelming they are in practice.
So if you’ve been worried about an AI apocalypse in cybersecurity, it’s best to keep your cool. The real threats are still coming from familiar places. And for now, AI’s bark is much louder than its byte.
