#ai generated malware

[the-wisper-report]

This Is My (Not) Surprised Face

Two days ago, Check Point Research published an article regarding what they’re calling the first evidence of a completely AI-driven malware framework, VoidLink. They first encountered this new framework in December, affecting Linux, and at the time considered the cluster of samples to be unfinished in terms of a fully functional tool. But it was evolving. Fast. Too fast for human oversight considering the level of sophistication and complexity.

The framework includes cloud implantation and harvesting credentials via Git repositories. It contains features like in-memory plugins, adaptive stealth capabilities tailored enough to preserve operational security over performance and multiple command-and-control channels. Earlier this month, CPR wrote about these details, concluding that it was likely inspired by Cobalt Strike’s Beacon Object Files approach. (For more information on that, see my report on Cobalt Strike here.) VoidLink is primarily written in Zig, but also has capability in Go, C and React. At the time of this earlier article, the malware was considered to be still in proof-of-concept mode, with no real intention of use being clear other than some kind of commercial venture.

Fast forward just a week, and we have our answer. VoidLink is a self-evolving, powerful malware suite, capable of building an entire attack campaign with little oversight. It is, however, not perfect. CPR has been monitoring the architecture traced back to a single human actor and found operational security (OPSEC) failures that allowed them to document how this malware is coded and how it evolves in real time. It is there that CPR concluded that this model was AI-generated, considering that what looked like a development period that should have taken many months was happening in a matter of hours or days.

Much of my research results in finding patterns. The ebb and flow of infosec is discernible if you spend enough time in it. And today, I’m having a moment of deeply conflicted vindication. I saw this coming, and have even reported on it several times, going back to September and October of 2025; PromptLock, prompt injection in general, and this report on generative AI that brings up novel malware forms. But it is not something I wanted to be right about. AI-generated code has evolved to be autonomous and continuous. CPR’s report highlights some of the structures that give away the AI nature of VoidLink, namely the thoroughness, speed, and adaptability in a modular framework.

The danger here isn’t actually the new and innovative ways AI can be manipulated by threat actors. That was always a foregone conclusion, since everything that exists is subject to exploitation (I mean, Cobalt Strike is a great example of that). It’s the sheer volume of it that is going to overwhelm attempts to mitigate it when fully functional frameworks can be built from the ground up in less than a week. And keeping up with it will be challenging for human SOC teams that require things like food and sleep. Reports like CPR’s will go a long way to understanding the infrastructure of AI malware, but that alone is not enough. We are currently living with what I consider an anti-vaxxer mentality towards cybersecurity. Too many companies either don’t understand why they need protection or don’t want to pay for it, or be compliant with it if they do. I predicted this trend months ago. Now it remains to be seen what we, as an industry, are going to do about it.

Posted on LinkedIn, 1/22/26

[the-wisper-report]

Digital Bumrush

Consider a scenario of a small group of defenders beset by overwhelming numbers. The cause looks hopeless, the attackers’ objective being to either starve them out or break through the defenses just one time to open the way for all out assault. It’s the setup of numerous action movies and fantasy stories. And in those stories, some fateful twist or last minute ingenuity saves the day and the defenders will win (or stalemate) somehow. Maybe reinforcements arrive. Maybe the walls hold and attrition whittles down the numbers. Maybe there’s mutually assured destruction. It’s a popular trope, and off the top of my head I can think of half a dozen movies that use it, with varying degrees of victory or lack thereof, for the hapless defenders.

We often say that art imitates life. But more often lately, life is imitating art. APT36, the Pakistan based threat actor also known as Transparent Tribe, has been using a campaign of AI-generated, middling to low quality but enormous quantity of malware attacks. Dubbed as ‘vibeware’, the emerging new family utilizes niche code languages like Nim, Zig, or Crystal to in essence copy the logic of more known languages like C and Go. And then a target is flooded with ‘disposable’ attacks in the hopes that one will slip past security simply through sheer volume. It’s less of an assault on defense as it is a tactic to overwhelm analysts trying to keep up with the noise. Bitedefender’s Radu Tudorica describes it as a distributed denial-of-detection.

The tactic takes advantage of something I’ve been concerned about for a while: the difference between human and machine speed. The trouble with AI-generated malware isn’t that it’s particularly sophisticated or innovative (although some of them are), it’s that it can evolve and shift at a rate that’s just impossible to keep up with for the real people trying to hold it at bay. Automation doesn’t suffer from fatigue or need lunch and bathroom breaks. In the hours when analysts are clocked out, the machine is still churning at scale. Regardless of my personal feelings on LLM’s, AI security is probably the best tool we have to fight against AI malware.

Vibeware often dead ends itself. Many versions of the generated code simply collapse under their own misconfigured or poorly written commands. But that isn’t the point of the attack strategy. Those sloppy, disposable bits are a feint. They make a lot of noise, filling up the logs with more traffic than one can reasonably be expected to sift through, and distract from the piece actually carrying a payload that’s now slipped by an exhausted triage analyst while they were busy tracking everything else. And even if the payload carrying malware is caught, there’s another one somewhere else also carrying a payload to another endpoint. All it takes is one successful incursion to become compromised; vibeware usually has several options, all aimed at a different aspect of the target, whether that’s the cloud services, the messaging apps, or the productivity ones.

So how do we stop it? ReversingLabs has also published a summary of vibeware, and quoted many experts in the security industry. Among them, Jason Soroko, a senior fellow at Sectigo, says that defending against automated malware is less about reacting to how it’s constructed and more about what it’s doing. Strictly enforcing zero-trust compliance to contain unauthorized outbound communication will cut down on the amount of ‘noise’ in the first place. Martin Zugec of Bitedefender says that organizations that have neglected their security hygiene are now going to have to play catch-up. Employing network segmentation, the principle of least privilege, and active endpoint monitoring, and making the environment hostile and unpredictable to attackers, he adds.

Ya know, the basics. I don’t what’s more astounding to me, that someone has figured out how to create digital siege engines or that we’re still having this conversation about baseline security in 2026. Either way, I expect this isn’t the only time I’ll be reporting on it.

Posted, 4/17/26