Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Max Heinemeyer
Global Field CISO
Share
01
Apr 2023
Email communications in the era of generative AI
Today’s headlines on cyber security attacks are dominated by serious breaches that cripple critical national infrastructure providers, wreak geopolitical havoc through nation state espionage, and create paralyzing financial demands through extortion and ransomware. But how do these breaches occur in the first place, and why do they keep happening?
Social engineering – specifically malicious cyber campaigns delivered via email - remain the primary source of an organization’s vulnerability to attack. A recent survey conducted by Darktrace to 6,700 global cross-industry professionals across the UK, US, Australia, France, Germany, and Netherlands, found almost a third (30%) have fallen victim to a fraudulent email. The risk, as ever, seems to be uncontrollably on the rise – 70% of respondents have noticed an increase in the frequency of malicious emails in the last six months.
At the end of last year, I wrote about how widespread accessibility to generative AI has compounded the problem, leading to likely efficiency gains for skilled threat actors launching email attacks. Just a few weeks ago, we published research which demonstrated that while the number of email attacks across our own customer base has remained steady since ChatGPT’s release, those that rely on tricking victims into clicking malicious links have declined, while linguistic complexity, including text volume, punctuation, and sentence length among others, have increased. This indicates that cyber-criminals may be redirecting their focus to crafting more sophisticated social engineering scams that exploit user trust delivered over email.
Our latest findings, conducted by Darktrace Research, today reveal a 135% increase in ‘novel social engineering attacks’ across thousands of active Darktrace/Email customers from January to February 2023, corresponding with the widespread adoption of ChatGPT. A novel social engineering phishing email is an email attack that shows a strong linguistic deviation – semantically and syntactically – compared to other phishing emails. The trend suggests that generative AI, such as ChatGPT, is providing an avenue for threat actors to craft sophisticated and targeted attacks at speed and scale.
Organizations are already well aware of the danger novel threats pose – globally, 82% of respondents reported are concerned that hackers can use generative AI to create scam emails that are indistinguishable from genuine communication.
Navigating the trust deficit
Employees are being exploited continuously through their primary means of communication - email. Given that humans, and our psychology, are at the heart of the problem, organizations have scrambled to develop awareness and training programs that are delivered routinely to employees in an attempt to help them defend themselves and their organization. While a defense-in-depth approach is generally a good practice to follow, the value of security awareness training, teaching employees to spot attempts at social engineering delivered in text format, is diminishing rapidly. This problem has become especially evident now that we find ourselves in a world where ‘bad’ emails look all but undistinguishable to the human eye from ‘good’ emails.
Over emphasizing the responsibility of employees poses a business challenge as well as a security problem: a bloated trust deficit. In times of economic uncertainty, business success is predicated on employee productivity, creativity, and efficiency. All three of these are squashed when employees treat the communication they receive, both internal and external, with forensic suspicion and mistrust.
As generative AI becomes more mainstream – across images, audio, video, and text – we anticipate that this erosion of trust in digital communication will only continue. What happens to the workplace if an employee is left questioning their own perception when communicating with their manager or a colleague, who they can see and hear over a video conference call, but who may be an entirely fictitious creation?
The way forward
The future lies in a partnership between artificial intelligence and human beings where the algorithms are responsible for determining whether communication is malicious or benign, thereby taking the burden of responsibility off the human. In turn, humans can pursue activities that drive productivity and performance. Current training and awareness programs, coupled with traditional email security technologies which assume that knowledge of past attacks will help predict campaigns, are not sustainable in this rapidly evolving landscape.
Our research reveals that traditional email security tools that rely on knowledge of past threats, take an average of thirteen days from an attack being launched on a victim to that attack being detected. This leaves defenders vulnerable for almost two weeks if they rely solely on traditional tools. When this statistic is contextualized against the 135% increase we’ve seen in novel social engineering campaigns, the inefficiency of ‘attack centric’ email security is made perfectly evident. If we don’t know how to define entirely new and unique attacks, how can we ever hope to recognize them based on past attacks?
A deep understanding of ‘you’ is needed in the face of these radical changes to the email threat landscape. Instead of trying to predict attacks, an understanding of your employees’ behavior, based from their email inbox, is needed to create patterns of life for every email user: their relationships, tone and sentiments and hundreds of other data points. By leveraging AI to combat email security threats, we not only reduce risk, but revitalize organizational trusts and contribute to business outcomes.
Ironically, generative AI may be worsening the social engineering challenge, but AI that knows you, could be the parry.
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Managing OT Remote Access with Zero Trust Control & AI Driven Detection
The shift toward IT-OT convergence
Recently, industrial environments have become more connected and dependent on external collaboration. As a result, truly air-gapped OT systems have become less of a reality, especially when working with OEM-managed assets, legacy equipment requiring remote diagnostics, or third-party integrators who routinely connect in.
This convergence, whether it’s driven by digital transformation mandates or operational efficiency goals, are making OT environments more connected, more automated, and more intertwined with IT systems. While this convergence opens new possibilities, it also exposes the environment to risks that traditional OT architectures were never designed to withstand.
The modernization gap and why visibility alone isn’t enough
The push toward modernization has introduced new technology into industrial environments, creating convergence between IT and OT environments, and resulting in a lack of visibility. However, regaining that visibility is just a starting point. Visibility only tells you what is connected, not how access should be governed. And this is where the divide between IT and OT becomes unavoidable.
Security strategies that work well in IT often fall short in OT, where even small missteps can lead to environmental risk, safety incidents, or costly disruptions. Add in mounting regulatory pressure to enforce secure access, enforce segmentation, and demonstrate accountability, and it becomes clear: visibility alone is no longer sufficient. What industrial environments need now is precision. They need control. And they need to implement both without interrupting operations. All this requires identity-based access controls, real-time session oversight, and continuous behavioral detection.
The risk of unmonitored remote access
This risk becomes most evident during critical moments, such as when an OEM needs urgent access to troubleshoot a malfunctioning asset.
Under that time pressure, access is often provisioned quickly with minimal verification, bypassing established processes. Once inside, there’s little to no real-time oversight of user actions whether they’re executing commands, changing configurations, or moving laterally across the network. These actions typically go unlogged or unnoticed until something breaks. At that point, teams are stuck piecing together fragmented logs or post-incident forensics, with no clear line of accountability.
In environments where uptime is critical and safety is non-negotiable, this level of uncertainty simply isn’t sustainable.
The visibility gap: Who’s doing what, and when?
The fundamental issue we encounter is the disconnect between who has access and what they are doing with it.
Traditional access management tools may validate credentials and restrict entry points, but they rarely provide real-time visibility into in-session activity. Even fewer can distinguish between expected vendor behavior and subtle signs of compromise, misuse or misconfiguration.
As a result, OT and security teams are often left blind to the most critical part of the puzzle, intent and behavior.
Closing the gaps with zero trust controls and AI‑driven detection
Managing remote access in OT is no longer just about granting a connection, it’s about enforcing strict access parameters while continuously monitoring for abnormal behavior. This requires a two-pronged approach: precision access control, and intelligent, real-time detection.
Zero Trust access controls provide the foundation.By enforcing identity-based, just-in-time permissions, OT environments can ensure that vendors and remote users only access the systems they’re explicitly authorized to interact with, and only for the time they need. These controls should be granular enough to limit access down to specific devices, commands, or functions. By applying these principles consistently across the Purdue Model, organizations can eliminate reliance on catch-all VPN tunnels, jump servers, and brittle firewall exceptions that expose the environment to excess risk.
Access control is only one part of the equation
Darktrace / OT complements zero trust controls with continuous, AI-driven behavioral detection. Rather than relying on static rules or pre-defined signatures, Darktrace uses Self-Learning AI to build a live, evolving understanding of what’s “normal” in the environment, across every device, protocol, and user. This enables real-time detection of subtle misconfigurations, credential misuse, or lateral movement as they happen, not after the fact.
By correlating user identity and session activity with behavioral analytics, Darktrace gives organizations the full picture: who accessed which system, what actions they performed, how those actions compared to historical norms, and whether any deviations occurred. It eliminates guesswork around remote access sessions and replaces it with clear, contextual insight.
Importantly, Darktrace distinguishes between operational noise and true cyber-relevant anomalies. Unlike other tools that lump everything, from CVE alerts to routine activity, into a single stream, Darktrace separates legitimate remote access behavior from potential misuse or abuse. This means organizations can both audit access from a compliance standpoint and be confident that if a session is ever exploited, the misuse will be surfaced as a high-fidelity, cyber-relevant alert. This approach serves as a compensating control, ensuring that even if access is overextended or misused, the behavior is still visible and actionable.
If a session deviates from learned baselines, such as an unusual command sequence, new lateral movement path, or activity outside of scheduled hours, Darktrace can flag it immediately. These insights can be used to trigger manual investigation or automated enforcement actions, such as access revocation or session isolation, depending on policy.
This layered approach enables real-time decision-making, supports uninterrupted operations, and delivers complete accountability for all remote activity, without slowing down critical work or disrupting industrial workflows.
Where Zero Trust Access Meets AI‑Driven Oversight:
Granular Access Enforcement: Role-based, just-in-time access that aligns with Zero Trust principles and meets compliance expectations.
Context-Enriched Threat Detection: Self-Learning AI detects anomalous OT behavior in real time and ties threats to access events and user activity.
Automated Session Oversight: Behavioral anomalies can trigger alerting or automated controls, reducing time-to-contain while preserving uptime.
Full Visibility Across Purdue Layers: Correlated data connects remote access events with device-level behavior, spanning IT and OT layers.
Scalable, Passive Monitoring: Passive behavioral learning enables coverage across legacy systems and air-gapped environments, no signatures, agents, or intrusive scans required.
Complete security without compromise
We no longer have to choose between operational agility and security control, or between visibility and simplicity. A Zero Trust approach, reinforced by real-time AI detection, enables secure remote access that is both permission-aware and behavior-aware, tailored to the realities of industrial operations and scalable across diverse environments.
Because when it comes to protecting critical infrastructure, access without detection is a risk and detection without access control is incomplete.
Product Marketing Manager, OT Security & Compliance
Blog
/
Network
/
November 20, 2025
Xillen Stealer Updates to Version 5 to Evade AI Detection
Introduction
Python-based information stealer “Xillen Stealer” has recently released versions 4 and 5, expanding its targeting and functionality. The cross-platform infostealer, originally reported by Cyfirma in September 2025, targets sensitive data including credentials, cryptocurrency wallets, system information, browser data and employs anti-analysis techniques.
The update to v4/v5 includes significantly more functionality, including:
Persistence
Ability to steal credentials from password managers, social media accounts, browser data (history, cookies and passwords) from over 100 browsers, cryptocurrency from over 70 wallets
Kubernetes configs and secrets
Docker scanning
Encryption
Polymorphism
System hooks
Peer-to-Peer (P2P) Command-and-Control (C2)
Single Sign-On (SSO) collector
Time-Based One-Time Passwords (TOTP) and biometric collection
EDR bypass
AI evasion
Interceptor for Two-Factor Authentication (2FA)
IoT scanning
Data exfiltration via Cloud APIs
Xillen Stealer is marketed on Telegram, with different licenses available for purchase. Users who deploy the malware have access to a professional-looking GUI that enables them to view exfiltrated data, logs, infections, configurations and subscription information.
Figure 1: Screenshot of the Xillen Stealer portal.
Technical analysis
The following technical analysis examines some of the interesting functions of Xillen Stealer v4 and v5. The main functionality of Xillen Stealer is to steal cryptocurrency, credentials, system information, and account information from a range of stores.
Xillen Stealer specifically targets the following wallets and browsers:
AITargetDectection
Figure 2: Screenshot of Xillen Stealer’s AI Target detection function.
The ‘AITargetDetection’ class is intended to use AI to detect high-value targets based on weighted indicators and relevant keywords defined in a dictionary. These indicators include “high value targets”, like cryptocurrency wallets, banking data, premium accounts, developer accounts, and business emails. Location indicators include high-value countries such as the United States, United Kingdom, Germany and Japan, along with cryptocurrency-friendly countries and financial hubs. Wealth indicators such as keywords like CEO, trader, investor and VIP have also been defined in a dictionary but are not in use at this time, pointing towards the group’s intent to develop further in the future.
While the class is named ‘AITargetDetection’ and includes placeholder functions for initializing and training a machine learning model, there is no actual implementation of machine learning. Instead, the system relies entirely on rule-based pattern matching for detection and scoring. Even though AI is not actually implemented in this code, it shows how malware developers could use AI in future malicious campaigns.
Figure 3: Screenshot of dead code function.
AI Evasion
Figure 4: Screenshot of AI evasion function to create entropy variance.
‘AIEvasionEngine’ is a module designed to help malware evade AI-based or behavior-based detection systems, such as EDRs and sandboxes. It mimics legitimate user and system behavior, injects statistical noise, randomizes execution patterns, and camouflages resource usage. Its goal is to make the malware appear benign to machine learning detectors. The techniques used to achieve this are:
Noise Injection: Performs random memory, CPU, file, and network operations to confuse behavioral classifiers
Timing Randomization: Introduces irregular delays and sleep patterns to avoid timing-based anomaly detection
Resource Camouflage: Adjusts CPU and memory usage to imitate normal apps (such as browsers, text editors)
API Call Obfuscation: Random system API calls and pattern changes to hide malicious intent
Memory Access Obfuscation: Alters access patterns and entropy to bypass ML models monitoring memory behavior
PolymorphicEngine
As part of the “Rust Engine” available in Xillen Stealer is the Polymorphic Engine. The ‘PolymorphicEngine’ struct implements a basic polymorphic transformation system designed for obfuscation and detection evasion. It uses predefined instruction substitutions, control-flow pattern replacements, and dead code injection to produce varied output. The mutate_code() method scans input bytes and replaces recognized instruction patterns with randomized alternatives, then applies control flow obfuscation and inserts non-functional code to increase variability. Additional features include string encryption via XOR and a stub-based packer.
Collectors
DevToolsCollector
Figure 5: Screenshot of Kubernetes data function.
The ‘DevToolsCollector’ is designed to collect sensitive data related to a wide range of developer tools and environments. This includes:
IDE configurations
VS Code, VS Code Insiders, Visual Studio
JetBrains: Intellij, PyCharm, WebStorm
Sublime
Atom
Notepad++
Eclipse
Cloud credentials and configurations
AWS
GCP
Azure
Digital Ocean
Heroku
SSH keys
Docker & Kubernetes configurations
Git credentials
Database connection information
HeidiSQL
Navicat
DBeaver
MySQL Workbench
pgAdmin
API keys from .env files
FTP configs
FileZilla
WinSCP
Core FTP
VPN configurations
OpenVPN
WireGuard
NordVPN
ExpressVPN
CyberGhost
Container persistence
Figure 6: Screenshot of Kubernetes inject function.
Biometric Collector
Figure 7: Screenshot of the ‘BiometricCollector’ function.
The ‘BiometricCollector’ attempts to collect biometric information from Windows systems by scanning the C:\Windows\System32\WinBioDatabase directory, which stores Windows Hello and other biometric configuration data. If accessible, it reads the contents of each file, encodes them in Base64, preparing them for later exfiltration. While the data here is typically encrypted by Windows, its collection indicates an attempt to extract sensitive biometric data.
Password Managers
The ‘PasswordManagerCollector’ function attempts to steal credentials stored in password managers including, OnePass, LastPass, BitWarden, Dashlane, NordPass and KeePass. However, this function is limited to Windows systems only.
SSOCollector
The ‘SSOCollector’ class is designed to collect authentication tokens related to SSO systems. It targets three main sources: Azure Active Directory tokens stored under TokenBroker\Cache, Kerberos tickets obtained through the klist command, and Google Cloud authentication data in user configuration folders. For each source, it checks known directories or commands, reads partial file contents, and stores the results as in a dictionary. Once again, this function is limited to Windows systems.
TOTP Collector
The ‘TOTP Collector’ class attempts to collect TOTPs from:
Authy Desktop by locating and reading from Authy.db SQLite databases
Microsoft Authenticator by scanning known application data paths for stored binary files
TOTP-related Chrome extensions by searching LevelDB files for identifiable keywords like “gauth” or “authenticator”.
Each method attempts to locate relevant files, parse or partially read their contents, and store them in a dictionary under labels like authy, microsoft_auth, or chrome_extension. However, as before, this is limited to Windows, and there is no handling for encrypted tokens.
Enterprise Collector
The ‘EnterpriseCollector’ class is used to extract credentials related to an enterprise Windows system. It targets configuration and credential data from:
The files and directories are located based on standard environment variables with their contents read in binary mode and then encoded in Base64.
Super Extended Application Collector
The ‘SuperExtendedApplication’ Collector class is designed to scan an environment for 160 different applications on a Windows system. It iterates through the paths of a wide range of software categories including messaging apps, cryptocurrency wallets, password managers, development tools, enterprise tools, gaming clients, and security products. The list includes but is not limited to Teams, Slack, Mattermost, Zoom, Google Meet, MS Office, Defender, Norton, McAfee, Steam, Twitch, VMWare, to name a few.
Bypass
AppBoundBypass
This code outlines a framework for bypassing App Bound protections, Google Chrome' s cookie encryption. The ‘AppBoundBypass’ class attempts several evasion techniques, including memory injection, dynamic-link library (DLL) hijacking, process hollowing, atom bombing, and process doppelgänging to impersonate or hijack browser processes. As of the time of writing, the code contains multiple placeholders, indicating that the code is still in development.
Steganography
The ‘SteganographyModule’ uses steganography (hiding data within an image) to hide the stolen data, staging it for exfiltration. Multiple methods are implemented, including:
Image steganography: LSB-based hiding
NTFS Alternate Data Streams
Windows Registry Keys
Slack space: Writing into unallocated disk cluster space
Polyglot files: Appending archive data to images
Image metadata: Embedding data in EXIF tags
Whitespace encoding: Hiding binary in trailing spaces of text files
Exfiltration
CloudProxy
Figure 8: Screenshot of the ‘CloudProxy’ class.
The CloudProxy class is designed for exfiltrating data by routing it through cloud service domains. It encodes the input data using Base64, attaches a timestamp and SHA-256 signature, and attempts to send this payload as a JSON object via HTTP POST requests to cloud URLs including AWS, GCP, and Azure, allowing the traffic to blend in. As of the time of writing, these public facing URLs do not accept POST requests, indicating that they are placeholders meant to be replaced with attacker-controlled cloud endpoints in a finalized build.
P2PEngine
Figure 9: Screenshot of the P2PEngine.
The ‘P2PEngine’ provides multiple methods of C2, including embedding instructions within blockchain transactions (such as Bitcoin OP_RETURN, Ethereum smart contracts), exfiltrating data via anonymizing networks like Tor and I2P, and storing payloads on IPFS (a distributed file system). It also supports domain generation algorithms (DGA) to create dynamic .onion addresses for evading detection.
After a compromise, the stealer creates both HTML and TXT reports containing the stolen data. It then sends these reports to the attacker’s designated Telegram account.
Xillen Killers
FIgure 10: Xillen Killers.
Xillen Stealer appears to be developed by a self-described 15-year-old “pentest specialist” “Beng/jaminButton” who creates TikTok videos showing basic exploits and open-source intelligence (OSINT) techniques. The group distributing the information stealer, known as “Xillen Killers”, claims to have 3,000 members. Additionally, the group claims to have been involved in:
Analysis of Project DDoSia, a tool reportedly used by the NoName057(16) group, revealing that rather functioning as a distributed denial-of-service (DDos) tool, it is actually a remote access trojan (RAT) and stealer, along with the identification of involved individuals.
Compromise of doxbin.net in October 2025.
Discovery of vulnerabilities on a Russian mods site and a Ukrainian news site
The group, which claims to be part of the Russian IT scene, use Telegram for logging, marketing, and support.
Conclusion
While some components of XillenStealer remain underdeveloped, the range of intended feature set, which includes credential harvesting, cryptocurrency theft, container targeting, and anti-analysis techniques, suggests that once fully developed it could become a sophisticated stealer. The intention to use AI to help improve targeting in malware campaigns, even though not yet implemented, indicates how threat actors are likely to incorporate AI into future campaigns. Credit to Tara Gould (Threat Research Lead)
