AI Wrote This Malware: Dissecting the Insides of a Vibe-Coded Malware Campaign

Authored by Aayush Tyagi  

Background 

The term ‘Vibe coding,’ first coined back in February of 2025 by OpenAI researchers, has exploded across digital platforms. With hundreds of articles and YouTube Videos discussing the dangers of Vibe coding and warning the internet about the rise of “Vibe Coders”, while others labelled it as the fundamental shift in software development and the future of coding.  

Vibe Coding is an approach where the AI does heavy lifting, rather than the user. Instead of manually writing code or implementing algorithms, users describe their intent through text-based prompt, and the LLMs respond with fully functional code and explanation. Unsurprisingly, the internet is now flooded with guides on the best LLMs and prompts to generate “perfect” code. 

Given the ease of generating fully functional code, McAfee Labs has also seen a rise in vibe-coded malware. In these campaigns, certain components of the kill chain contain AI-generated code, significantly reducing the effort and knowledge required to execute new malware campaigns. This shift not only makes malware campaigns more scalable but also lowers the barrier to entry for new malware authors. 

Executive summary 

In January 2026, McAfee Labs observed 443 malicious zip files impersonating a wide range of software, including AI image generators and voice-changing tools, stock-market trading utilities, game mods and modding tools, game hacks, graphics card and USB drivers, ransomware decryptors, VPNs, emulators, and even infostealer, cookie-stealer, and backdoor malware, to infect users.  

Across the 440+ zip files, we observed 48 unique malicious WinUpdateHelper.dll variants, responsible for the infections. McAfee has been detecting variants of this threat since December 2024, although the vibe coding observed in certain components appears to be a recent addition. These files are distributed through various legitimate content delivery network (CDN) services and file-hosting websites, such as Discord, SourceForge, FOSSHub, and MediaFire, to name a few. Another website that was actively delivering this malware was mydofiles[.]com. 

Here, the attackers implement volume-driven malware distribution techniques to infect as many users as possible.  

This attack begins when users surf the internet looking for tools and software that promise to simplify their tasks. Instead, they encounter trojanized zip files.  

We discovered over 100 URLs actively spreading this malware, of which approximately 61 were hosted on Discord, 17 on SourceForge, and 15 on mydofiles[.]com. 

On running the executable, it loads a malicious WinUpdateHelper.dll file, which redirects the user to file-hosting websites, under the disguise that they are missing crucial dependencies and tricks them into installing unrelated software, which is a distraction. Meanwhile, the DLL has already requested and executed a malicious PowerShell script from a command-and-control (C2) server.  

This script infects the user’s system and downloads additional mining software, and abuses the system’s resources, or it downloads additional payloads such as SalatStealer or Mesh Agent, depending on the WinUpdateHelper.dll sample which infected the user.  

In this PowerShell script, the presence of explanatory comments and structured sections strongly indicates the use of LLM models to generate this code. 

Read more about this in the Using AI to generate malware? section below.  

So far, we’ve observed the mining of Ravencoin, Zephyr, Monero, Bitcoin Gold, Ergo, and Clore cryptocurrencies.    

Due to the presence of hardcoded Bitcoin wallet credentials within these malware samples, we were able to trace on-chain transactions and identify wallets containing over $4,500 USD that are part of this campaign.  

Since most of the mining activity targets privacy-focused cryptocurrencies such as Zephyr, Ravencoin and Monero, the real financial impact is likely to be nearly double the amount identified through Bitcoin tracing alone.  

Geographical Prevalence 

This malware campaign has specifically targeted users in the following counties, ranked by prevalence: The United States of America, followed by United Kingdom, India, Brazil, France, Canada, Australia. 

Bottom Line

The availability of LLMs capable of generating code instantly, combined with the widespread accessibility of technical knowledge, has created a low-effort, high-reward environment, making malware deployment increasingly accessible. 

At McAfee Labs, we have been doing hard work so that you don’t need to worry. But it always helps to be informed and educated on the latest threat that steps into the threat landscape. 
We will continue monitoring these campaigns to ensure our customers remain informed and protected across platforms. 

Technical Analysis  

Impersonated Applications

Here we see malware distribution at a large scale and by analyzing the filenames of these ZIP archives, we can infer to the users that are being targeted. These are some of the names we’ve witnessed in the wild. 

The attackers are actively impersonating video game cheats and game mods for popular titles, and well-known script executors for Roblox, such as Delta Executor and Solara as seen above.  

Names such as Panther-Stealer and Zerotrace-Stealer indicate that even users looking for malware on the internet are not safe either, reinforcing the notion that there is truly no honor among thieves. 

The campaign also leverages drivers and AI-themed tools as part of its lure portfolio among other tools. Interestingly, we see the name ‘DeepSeek.zip’, where attackers are exploiting a prominent LLM model, DeepSeek. McAfee had encountered these types of attacks in early 2025 and covered them extensively.  

Read the previous blog here: Look Before You Leap: Imposter DeepSeek Software Seek Gullible Users  

Stage 1 Payload – Misleading Installation  

Once the user downloads the ZIP archive from Discord or any other website. They get the following set of files.

Here, the executable named ‘gta-5-online-mod-menu.exe’ (Highlighted in Blue) is a legitimate and clean file. Whereas the file named ‘WinUpdateHelper.dll’ (Highlighted in Red) is malicious.  

On executing ‘gta-5-online-mod-menu.exe’, the malicious DLL is loaded. The user is informed that they are missing dependencies, and they’re redirected to the following URL via default browser.  

hxxps://igk[.]filexspace.com/getfile/XKQLPSK?title=DependencyCore&tracker=gta-5-online-mod-menu 

Here, within the URL, a tracker variable is used to identify which malware has infected the user. In this instance, it was ‘gta-5-online-mod-menu’.  

Dependecycore.zip is a setup file. On execution, it installs unrelated 3rd party software on the victim’s system. 

In this instance, iTop Easy Desktop was installed. 

This unwanted installation is meant to subvert users’ attention. As, the WinUpdateHelper.dll has already connected to the C2 server and infected the system.   

Stage 1 Payload – Malicious Functionality  

Once the redirection code is executed, the malware executes the malicious code.  

In the above code snippet, which is present in the WinUpdateHelper.dll, we can see that a new service has been created under the name “Microsoft Console Host” to make it appear to be benign (Highlighted in Red). The parameters passed to this service ensure that it executes at system boot. This is done to maintain persistence in the system.

The service executes a PowerShell command that dynamically generates the C2 domain using the UNIX time stamp.  

Using the following code, 
$([Math]::Floor([DateTimeOffset]::UtcNow.ToUnixTimeSeconds() / 5000000) * 5000000).xyz 

It generates a domain name that changes once every 5,000,000 seconds or 58 days. 

The latest C2 domain we’ve discovered that is up and running is 
1770000000[.]xyz/script?id=fA9zQk2L0M&tag=WinUpdateHelper

During our analysis we observed the following domain 
1765000000[.]xyz/script?id=fA9zQk2L0M&tag=WinUpdateHelper, which is present in the following images.  

Here the id=fA9zQk2L0M is randomly generated, to uniquely identify the user and tag=WinUpdateHelper is used to identify the malware campaign.  

The malware connects to the above-mentioned C2 server to download a PowerShell script and execute it in memory. This fileless execution ensures improved evasion against signature-based detections. 

Stage 2 Payload – PowerShell Script  

It is funny to note here, that the first comment of this script says “# I am forever sorry” which indicates that the attacks do carry some guilt regarding their actions, but not enough to stop the campaign. We found similar comments, such as “# sorry lol”, across multiple PowerShell scripts we discovered.  

The first set of commands (Highlighted in Green) are used to delete windows services and scheduled tasks. This is done to remove older or conflicting persistence mechanisms and to avoid duplicate miners from running on the same system. 

The second set of commands (Highlighted in Red) are registry modifications, that adds “C:\ProgramData” to Windows Defender exclusion paths. That is, ProgramData Folder won’t be scanned by Windows Defender anymore. This exclusion allows malware to drop additional payloads to disk, without the risk of them being detected and removed.  

The third set of commands (Highlighted in Blue) does exactly that. It downloads the next level payload from the URL “hxxps://1765000000[.]xyz/download/xbhgjahddaa” and stored it at this path “C:\ProgramData\fontdrvhost.exe”.

Again the name ‘fontdrvhost.exe’ imitates a legitimate Windows binary, to masquerade its true intent. After the download, the file is decoded using a simple arithmetic decryption routine. This provides protection against static signature detection and network detection. 

The payload is an XMRIG miner sample. In the next command, the miner is initialized and executed. Here, we see the miner connecting to “solo-zeph.2miners.com:4444” and start CPU based Zephyr coin mining using the following wallet address: ‘ZEPHsCY4zbcHGgz2U8PvkEjkWjopuPurPNv8nnSFnM5MN8hBas8kBN4hoNKmc7uMRfUQh4Fc9AHyGxL6NFARnc217m2vYgbKxf’. 

In the second half of the script, we see another miner being set up and executed using the same technique (Highlighted in Red). This time the file is stored as “RuntimeBroker.exe” in the ProgramData folder. The miner is connecting to “solo-rvn.2miners.com:7070” to mine Ravencoin and it is using the system’s GPU instead of the CPU for mining (Highlighted in Blue).  

This is the wallet address used for mining in this instance ‘bc1q9a59scnfwkdlm6wlcu5w76zm2uesjrqdy4fr8r’.  

Hence, we see a dual coin-mining deployment infrastructure utilizing both CPU and GPU resources to optimize mining efficiency. 

Bitcoin? Interesting…  

What is interesting here is that attackers have used a bitcoin wallet address for mining Ravencoin, which indicates they are using multi-coin pools for mining. The attackers are using the victims’ machine to mine Ravencoin and automatically convert the mining rewards to Bitcoin before the payout.  

This is done for a variety of reasons, such as, bitcoin offers higher liquidity and has broader acceptance, but most importantly, Ravencoin is computationally easier and economically viable to mine on victim’s system. Bitcoin requires specialized ASIC hardware for profitable mining and attempting to mine Bitcoin directly on infected systems would generate negligible returns. We’ve seen the same behaviour in multiple samples. 

This is a smoking gun. Unlike Zephyr coin or Monero, Bitcoin’s blockchain is fully traceable. Every Satoshi, the smallest unit of Bitcoin, can be traced across the blockchain from the moment it was mined to its current holder. From there, it becomes easy to determine how much cryptocurrency the threat actor is receiving. More on this later.  

Anti-Analysis Techniques 

The attackers have meticulously designed the campaign and have implemented various anti-analysis techniques to thwart researchers.  

The PowerShell script we’ve seen above is responsible for downloading and initializing the coin miner samples. It is only accessible via PowerShell. If we try to access the server via Curl, we get the following response.  

 This indicates that the server is actively monitoring the User-Agent of incoming requests and deploys the payload only when the request originates from PowerShell. 

 Similarly, the URLs embedded within the PowerShell script that download the next payload are unique to each victim and remain active for 60 seconds. After that, they return a 404 Not Found error.  

These techniques are meant to confuse and disorient researchers, making the analysis difficult.  

Using AI to generate malware?  

While working on this malware campaign, we came across over 440 unique zip files. These same zip files were distributed with over 1700 different names, targeting various software. 

Across these 440 zip files, we noticed 48 unique variants of WinUpdateHelper.dll. These 48 files can be clustered together into 17 distinct kill chains, each featuring their own C2 infrastructure, misleading installation setups, second-stage PowerShell scripts and final payloads, yet the cryptocurrency wallet credentials remain similar. 

In the above technical analysis, we’ve only covered 1 kill chain. Yet, across these 17 kill chains, we’ve noticed the flow remain the same.  

Across multiple second stage payloads, we encounter multiple comments such as the following, embedded within the code:

# === Create and execute run.bat in C:\ProgramData ===

:: This batch file:

:: – Creates the hidden folder C:\ProgramData\cvtres if it doesn”t exist (using CMD attrib for hidden + system)

:: – Downloads cvtres.exe from your GitHub URL

:: – Saves it to C:\ProgramData\cvtres\cvtres.exe

:: – Executes it immediately

:: – Runs completely hidden/minimized (no window visible)

The presence of such explanatory-style comments indicates that large language models were likely used during the development of these scripts. Especially, the comment “Downloads cvtres.exe from your GitHub URL”, where ‘Your GitHub URL’ refers to the threat actor’s GitHub repository that is hosting the malware, which indicates potential vibe coding.  

Tracking Bitcoin Across the Blockchain 

During analysis of this malware campaign, we came across few instances where the final payload was Infostealer malware. In most cases it was coin miner samples. 
In these cases, we encountered wallet credentials and mining pool URLs for several alternative cryptocurrencies such as Ravencoin, Zephyr, Monero, which aren’t traceable.  

Fortunately, we came across 7 bitcoin wallets that are part of this malware campaign and are actively receiving mined cryptocurrency. 

bc1q9a59scnfwkdlm6wlcu5w76zm2uesjrqdy4fr8r     bc1q7cpwxjatrtpa29u85tayvggs67f6fxwyggm8kd 

bc1qyy0cv8snz7zqummg0yucdfzpxv2a5syu7xzsdq    bc1qxhp6mn0h7k9r89w8amalqjn38t4j5yaa7t89rp 

bc1qxnkkpnuhydckmpx8fmkp73e38dfed93uhfh68l    bc1qrtztxnqnjk9q4d5hupnla245c7620ncj3tzp7h 

bc1q97yd574m9znar99fa0u799rvm55tnjzkw9l33w 

As of writing this blog, these wallets contain Bitcoin valued at approximately $4,536.20 USD. 

These wallets have seen regular withdrawals, with total funds received amounting to approximately $11,497.7 USD. 

McAfee Coverage 

McAfee has extensive coverage for this Coinminer Malware Campaign. We’re proactively covering new samples observed in the wild. 

Trojan:Win/Phishing.AP 

Trojan:Script/Coinminer.AT 

Trojan:Win/Dropper.AT 

Indicator of Compromise(s)