Black Hat researcher shows AI-powered malware can slip past Microsoft Defender
At the upcoming Black Hat security conference in Las Vegas, a new AI-powered malware tool is set to make waves—not because it’s the most dangerous we’ve ever seen, but because it demonstrates a clear and troubling evolution in cyber threats. According to reporting from Dark Reading and Tom’s Hardware, Kyle Avery from Outflank will present a lightweight, open-source language model engineered specifically to evade Microsoft Defender—the default security system used by over a billion Windows 10 and Windows 11 machines worldwide.
At first glance, an 8 percent evasion rate might seem unimpressive. But in cybersecurity terms, that’s an alarming leap. It means the AI model can bypass Defender in nearly one out of every 12 attempts—without relying on exotic attack vectors or massive infrastructure. When scaled across Windows’ 1.4 billion active user base, such a rate could translate to tens of millions of compromised machines, especially if the attack method were mass-deployed through AI-powered automation or plug-and-play malware kits.
What’s especially concerning is the model’s accessibility. Avery reportedly trained it for just $1,500 over three months using public tools and datasets. That cost is trivial in the world of cybercrime, suggesting that similar models could easily proliferate among threat actors. With enough scale and speed, even small gains in Defender evasion can be multiplied into significant systemic risks.
Fortunately, Avery is debuting his work in a responsible disclosure context, giving Microsoft and other defenders time to study the findings and adapt. But this is another stark reminder that generative AI isn’t just a novelty or productivity enhancer—it’s rapidly becoming a core tool in the offensive playbook for cybercriminals. Microsoft, already weathering layoffs in its security departments, now faces the pressing challenge of staying ahead of threats that are learning and evolving on their own.
