Blog
/
AI
/
October 30, 2023

Exploring AI Threats: Package Hallucination Attacks

Learn how malicious actors exploit errors in generative AI tools to launch packet attacks. Read how Darktrace products detect and prevent these threats!
Inside the SOC
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
Charlotte Thompson
Cyber Analyst
Written by
Tiana Kelly
Deputy Team Lead, London & Cyber Analyst
Default blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog image
30
Oct 2023

AI tools open doors for threat actors

On November 30, 2022, the free conversational language generation model ChatGPT was launched by OpenAI, an artificial intelligence (AI) research and development company. The launch of ChatGPT was the culmination of development ongoing since 2018 and represented the latest innovation in the ongoing generative AI boom and made the use of generative AI tools accessible to the general population for the first time.

ChatGPT is estimated to currently have at least 100 million users, and in August 2023 the site reached 1.43 billion visits [1]. Darktrace data indicated that, as of March 2023, 74% of active customer environments have employees using generative AI tools in the workplace [2].

However, with new tools come new opportunities for threat actors to exploit and use them maliciously, expanding their arsenal.

Much consideration has been given to mitigating the impacts of the increased linguistic complexity in social engineering and phishing attacks resulting from generative AI tool use, with Darktrace observing 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 and its peers [3].

Less overall consideration, however, has been given to impacts stemming from errors intrinsic to generative AI tools. One of these errors is AI hallucinations.

What is an AI hallucination?

AI “hallucination” is a term which refers to the predictive elements of generative AI and LLMs’ AI model gives an unexpected or factually incorrect response which does not align with its machine learning training data [4]. This differs from regular and intended behavior for an AI model, which should provide a response based on the data it was trained upon.  

Why are AI hallucinations a problem?

Despite the term indicating it might be a rare phenomenon, hallucinations are far more likely than accurate or factual results as the AI models used in LLMs are merely predictive and focus on the most probable text or outcome, rather than factual accuracy.

Given the widespread use of generative AI tools in the workplace employees are becoming significantly more likely to encounter an AI hallucination. Furthermore, if these fabricated hallucination responses are taken at face value, they could cause significant issues for an organization.

Use of generative AI in software development

Software developers may use generative AI for recommendations on how to optimize their scripts or code, or to find packages to import into their code for various uses. Software developers may ask LLMs for recommendations on specific pieces of code or how to solve a specific problem, which will likely lead to a third-party package. It is possible that packages recommended by generative AI tools could represent AI hallucinations and the packages may not have been published, or, more accurately, the packages may not have been published prior to the date at which the training data for the model halts. If these hallucinations result in common suggestions of a non-existent package, and the developer copies the code snippet wholesale, this may leave the exchanges vulnerable to attack.

Research conducted by Vulcan revealed the prevalence of AI hallucinations when ChatGPT is asked questions related to coding. After sourcing a sample of commonly asked coding questions from Stack Overflow, a question-and-answer website for programmers, researchers queried ChatGPT (in the context of Node.js and Python) and reviewed its responses. In 20% of the responses provided by ChatGPT pertaining to Node.js at least one un-published package was included, whilst the figure sat at around 35% for Python [4].

Hallucinations can be unpredictable, but would-be attackers are able to find packages to create by asking generative AI tools generic questions and checking whether the suggested packages exist already. As such, attacks using this vector are unlikely to target specific organizations, instead posing more of a widespread threat to users of generative AI tools.

Malicious packages as attack vectors

Although AI hallucinations can be unpredictable, and responses given by generative AI tools may not always be consistent, malicious actors are able to discover AI hallucinations by adopting the approach used by Vulcan. This allows hallucinated packages to be used as attack vectors. Once a malicious actor has discovered a hallucination of an un-published package, they are able to create a package with the same name and include a malicious payload, before publishing it. This is known as a malicious package.

Malicious packages could also be recommended by generative AI tools in the form of pre-existing packages. A user may be recommended a package that had previously been confirmed to contain malicious content, or a package that is no longer maintained and, therefore, is more vulnerable to hijack by malicious actors.

In such scenarios it is not necessary to manipulate the training data (data poisoning) to achieve the desired outcome for the malicious actor, thus a complex and time-consuming attack phase can easily be bypassed.

An unsuspecting software developer may incorporate a malicious package into their code, rendering it harmful. Deployment of this code could then result in compromise and escalation into a full-blown cyber-attack.

Figure 1: Flow diagram depicting the initial stages of an AI Package Hallucination Attack.

For providers of Software-as-a-Service (SaaS) products, this attack vector may represent an even greater risk. Such organizations may have a higher proportion of employed software developers than other organizations of comparable size. A threat actor, therefore, could utilize this attack vector as part of a supply chain attack, whereby a malicious payload becomes incorporated into trusted software and is then distributed to multiple customers. This type of attack could have severe consequences including data loss, the downtime of critical systems, and reputational damage.

How could Darktrace detect an AI Package Hallucination Attack?

In June 2023, Darktrace introduced a range of DETECT™ and RESPOND™ models designed to identify the use of generative AI tools within customer environments, and to autonomously perform inhibitive actions in response to such detections. These models will trigger based on connections to endpoints associated with generative AI tools, as such, Darktrace’s detection of an AI Package Hallucination Attack would likely begin with the breaching of one of the following DETECT models:

  • Compliance / Anomalous Upload to Generative AI
  • Compliance / Beaconing to Rare Generative AI and Generative AI
  • Compliance / Generative AI

Should generative AI tool use not be permitted by an organization, the Darktrace RESPOND model ‘Antigena / Network / Compliance / Antigena Generative AI Block’ can be activated to autonomously block connections to endpoints associated with generative AI, thus preventing an AI Package Hallucination attack before it can take hold.

Once a malicious package has been recommended, it may be downloaded from GitHub, a platform and cloud-based service used to store and manage code. Darktrace DETECT is able to identify when a device has performed a download from an open-source repository such as GitHub using the following models:

  • Device / Anomalous GitHub Download
  • Device / Anomalous Script Download Followed By Additional Packages

Whatever goal the malicious package has been designed to fulfil will determine the next stages of the attack. Due to their highly flexible nature, AI package hallucinations could be used as an attack vector to deliver a large variety of different malware types.

As GitHub is a commonly used service by software developers and IT professionals alike, traditional security tools may not alert customer security teams to such GitHub downloads, meaning malicious downloads may go undetected. Darktrace’s anomaly-based approach to threat detection, however, enables it to recognize subtle deviations in a device’s pre-established pattern of life which may be indicative of an emerging attack.

Subsequent anomalous activity representing the possible progression of the kill chain as part of an AI Package Hallucination Attack could then trigger an Enhanced Monitoring model. Enhanced Monitoring models are high-fidelity indicators of potential malicious activity that are investigated by the Darktrace analyst team as part of the Proactive Threat Notification (PTN) service offered by the Darktrace Security Operation Center (SOC).

Conclusion

Employees are often considered the first line of defense in cyber security; this is particularly true in the face of an AI Package Hallucination Attack.

As the use of generative AI becomes more accessible and an increasingly prevalent tool in an attacker’s toolbox, organizations will benefit from implementing company-wide policies to define expectations surrounding the use of such tools. It is simple, yet critical, for example, for employees to fact check responses provided to them by generative AI tools. All packages recommended by generative AI should also be checked by reviewing non-generated data from either external third-party or internal sources. It is also good practice to adopt caution when downloading packages with very few downloads as it could indicate the package is untrustworthy or malicious.

As of September 2023, ChatGPT Plus and Enterprise users were able to use the tool to browse the internet, expanding the data ChatGPT can access beyond the previous training data cut-off of September 2021 [5]. This feature will be expanded to all users soon [6]. ChatGPT providing up-to-date responses could prompt the evolution of this attack vector, allowing attackers to publish malicious packages which could subsequently be recommended by ChatGPT.

It is inevitable that a greater embrace of AI tools in the workplace will be seen in the coming years as the AI technology advances and existing tools become less novel and more familiar. By fighting fire with fire, using AI technology to identify AI usage, Darktrace is uniquely placed to detect and take preventative action against malicious actors capitalizing on the AI boom.

Credit to Charlotte Thompson, Cyber Analyst, Tiana Kelly, Analyst Team Lead, London, Cyber Analyst

References

[1] https://seo.ai/blog/chatgpt-user-statistics-facts

[2] https://darktrace.com/news/darktrace-addresses-generative-ai-concerns

[3] https://darktrace.com/news/darktrace-email-defends-organizations-against-evolving-cyber-threat-landscape

[4] https://vulcan.io/blog/ai-hallucinations-package-risk?nab=1&utm_referrer=https%3A%2F%2Fwww.google.com%2F

[5] https://twitter.com/OpenAI/status/1707077710047216095

[6] https://www.reuters.com/technology/openai-says-chatgpt-can-now-browse-internet-2023-09-27/

Inside the SOC
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
Charlotte Thompson
Cyber Analyst
Written by
Tiana Kelly
Deputy Team Lead, London & Cyber Analyst

More in this series

No items found.

Blog

/

Cloud

/

July 10, 2025

Crypto Wallets Continue to be Drained in Elaborate Social Media Scam

password on computer screenDefault blog imageDefault blog image

Overview

Continued research by Darktrace has revealed that cryptocurrency users are being targeted by threat actors in an elaborate social engineering scheme that continues to evolve. In December 2024, Cado Security Labs detailed a campaign targeting Web 3 employees in the Meeten campaign. The campaign included threat actors setting up meeting software companies to trick users into joining meetings and installing the information stealer Realst disguised as video meeting software.

The latest research from Darktrace shows that this campaign is still ongoing and continues to trick targets to download software to drain crypto wallets. The campaign features:

  • Threat actors creating fake startup companies with AI, gaming, video meeting software, web3 and social media themes.
  • Use of compromised X (formerly Twitter) accounts for the companies and employees - typically with verification to contact victims and create a facade of a legitimate company.
  • Notion, Medium, Github used to provide whitepapers, project roadmaps and employee details.
  • Windows and macOS versions.
  • Stolen software signing certificates in Windows versions for credibility and defense evasion.
  • Anti-analysis techniques including obfuscation, and anti-sandboxing.

To trick as many victims as possible, threat actors try to make the companies look as legitimate as possible. To achieve this, they make use of sites that are used frequently with software companies such as Twitter, Medium, Github and Notion. Each company has a professional looking website that includes employees, product blogs, whitepapers and roadmaps. X is heavily used to contact victims, and to increase the appearance of legitimacy. Some of the observed X accounts appear to be compromised accounts that typically are verified and have a higher number of followers and following, adding to the appearance of a real company.

Example of a compromised X account to create a “BuzzuAI” employee.
Figure 1: Example of a compromised X account to create a “BuzzuAI” employee.

The threat actors are active on these accounts while the campaign is active, posting about developments in the software, and product marketing. One of the fake companies part of this campaign, “Eternal Decay”, a blockchain-powered game, has created fake pictures pretending to be presenting at conferences to post on social media, while the actual game doesn’t exist.

From the Eternal Decay X account, threat actors have altered a photo from an Italian exhibition (original on the right) to make it look like Eternal Decay was presented.
Figure 2: From the Eternal Decay X account, threat actors have altered a photo from an Italian exhibition (original on the right) to make it look like Eternal Decay was presented.

In addition to X, Medium is used to post blogs about the software. Notion has been used in various campaigns with product roadmap details, as well as employee lists.

Notion project team page for Swox.
Figure 3: Notion project team page for Swox.

Github has been used to detail technical aspects of the software, along with Git repositories containing stolen open-source projects with the name changed in order to make the code look unique. In the Eternal Decay example, Gitbook is used to detail company and software information. The threat actors even include company registration information from Companies House, however they have linked to a company with a similar name and are not a real registered company.

 From the Eternal Decay Gitbook linking to a company with a similar name on Companies House.
Figure 4: From the Eternal Decay Gitbook linking to a company with a similar name on Companies House.
Gitbook for “Eternal Decay” listing investors.
Figure 5: Gitbook for “Eternal Decay” listing investors.
Gameplay images are stolen from a different game “Zombie Within” and posted pretending to be Eternal Decay gameplay.
Figure 6: Gameplay images are stolen from a different game “Zombie Within” and posted pretending to be Eternal Decay gameplay.

In some of the fake companies, fake merchandise stores have even been set up. With all these elements combined, the threat actors manage to create the appearance of a legitimate start-up company, increasing their chances of infection.

Each campaign typically starts with a victim being contacted through X messages, Telegram or Discord. A fake employee of the company will contact a victim asking to test out their software in exchange for a cryptocurrency payment. The victim will be directed to the company website download page, where they need to enter a registration code, provided by the employee to download a binary. Depending on their operating system, the victim will be instructed to download a macOS DMG (if available) or a Windows Electron application.

Example of threat actor messaging a victim on X with a registration code.
Figure 7: Example of threat actor messaging a victim on X with a registration code.

Windows Version

Similar to the aforementioned Meeten campaign, the Windows version being distributed by the fake software companies is an Electron application. Electron is an open-source framework used to run Javascript apps as a desktop application. Once the user follows directions sent to them via message, opening the application will bring up a Cloudflare verification screen.

Cloudflare verification screen.
Figure 8: Cloudflare verification screen.

The malware begins by profiling the system, gathering information like the username, CPU and core count, RAM, operating system, MAC address, graphics card, and UUID.

Code from the Electron app showing console output of system profiling.
Figure 9: Code from the Electron app showing console output of system profiling.

A verification process occurs with a captcha token extracted from the app-launcher URL and sent along with the system info and UUID. If the verification is successful, an executable or MSI file is downloaded and executed quietly. Python is also retrieved and stored in /AppData/Temp, with Python commands being sent from the command-and-control (C2) infrastructure.

Code from the Electron app looping through Python objects.
Figure 10: Code from the Electron app looping through Python objects.

As there was no valid token, this process did not succeed. However, based on previous campaigns and reports from victims on social media, an information stealer targeting crypto wallets is executed at this stage. A common tactic in the observed campaigns is the use of stolen code signing certificates to evade detection and increase the appearance of legitimate software. The certificates of two legitimate companies Jiangyin Fengyuan Electronics Co., Ltd. and Paperbucketmdb ApS (revoked as of June 2025) were used during this campaign.

MacOS Version

For companies that have a macOS version of the malware, the user is directed to download a DMG. The DMG contains a bash script and a multiarch macOS binary. The bash script is obfuscated with junk, base64 and is XOR’d.

Obfuscated Bash script.
Figure 11: Obfuscated Bash script.

After decoding, the contents of the script are revealed showing that AppleScript is being used. The script looks for disk drives, specifically for the mounted DMG “SwoxApp” and moves the hidden .SwoxApp binary to /tmp/ and makes it executable. This type of AppleScript is commonly used in macOS malware, such as Atomic Stealer.

AppleScript used to mount the malware and make it executable.
Figure 12: AppleScript used to mount the malware and make it executable.

The SwoxApp binary is the prominent macOS information stealer Atomic Stealer. Once executed the malware performs anti-analysis checks for QEMU, VMWare and Docker-OSX, the script exits if these return true.  The main functionality of Atomic Stealer is to steal data from stores including browser data, crypto wallets, cookies and documents. This data is compressed into /tmp/out.zip and sent via POST request to 45[.]94[.]47[.]167/contact. An additional bash script is retrieved from 77[.]73[.]129[.]18:80/install.sh.

Additional Bash script ”install.sh”.
Figure 13: Additional Bash script ”install.sh”.

Install.sh, as shown in Figure 13, retrieves another script install_dynamic.sh from the server https://mrajhhosdoahjsd[.]com. Install_dynamic.sh downloads and extracts InstallerHelper.app, then sets up persistence via Launch Agent to run at login.

Persistence added via Plist configuration.
Figure 14: Persistence added via Plist configuration.

This plist configuration installs a macOS LaunchAgent that silently runs the app at user login. RunAtLoad and KeepAlive keys are used to ensure the app starts automatically and remains persistent.

The retrieved binary InstallerHelper is an Objective-C/Swift binary that logs active application usage, window information, and user interaction timestamps. This data is written to local log files and periodically transmits the contents to https://mrajhhoshoahjsd[.]com/collect-metrics using scheduled network requests.

List of known companies

Darktrace has identified a number of the fake companies used in this scam. These can be found in the list below:

Pollens AI
X: @pollensapp, @Pollens_app
Website: pollens.app, pollens.io, pollens.tech
Windows: 02a5b35be82c59c55322d2800b0b8ccc
Notes: Posing as an AI software company with a focus on “collaborative creation”.

Buzzu
X: @BuzzuApp, @AI_Buzzu, @AppBuzzu, @BuzzuApp
Website: Buzzu.app, Buzzu.us, buzzu.me, Buzzu.space
Windows: 7d70a7e5661f9593568c64938e06a11a
Mac: be0e3e1e9a3fda76a77e8c5743dd2ced
Notes: Same as Pollens including logo but with a different name.

Cloudsign
X: @cloudsignapp
Windows: 3a3b13de4406d1ac13861018d74bf4b2
Notes: Claims to be a document signing platform.

Swox
X: @SwoxApp, @Swox_AI, @swox_app, @App_Swox, @AppSwox, @SwoxProject, @ProjectSwox
Website: swox.io, swox.app, swox.cc, swoxAI.com, swox.us
Windows: d50393ba7d63e92d23ec7d15716c7be6
Mac: 81996a20cfa56077a3bb69487cc58405ced79629d0c09c94fb21ba7e5f1a24c9
Notes: Claims to be a “Next gen social network in the WEB3”. Same GitHub code as Pollens.

KlastAI
X: Links to Pollens X account
Website: Links to pollens.tech
Notes: Same as Pollens, still shows their branding on its GitHub readme page.

Wasper
X: @wasperAI, @WasperSpace
Website: wasper.pro, wasper.app, wasper.org, wasper.space
Notes: Same logo and GitHub code as Pollens.

Lunelior
Website: lunelior.net, Lunelior.app, lunelior.io, lunelior.us
Windows: 74654e6e5f57a028ee70f015ef3a44a4
Mac: d723162f9197f7a548ca94802df74101

BeeSync
X: @BeeSyncAI, @AIBeeSync
Website: beesync.ai, beesync.cc
Notes: Previous alias of Buzzu, Git repo renamed January 2025.

Slax
X: @SlaxApp, @Slax_app, @slaxproject
Website: slax.tech, slax.cc, slax.social, slaxai.app

Solune
X: @soluneapp
Website: solune.io, solune.me
Windows: 22b2ea96be9d65006148ecbb6979eccc

Eternal Decay
X: @metaversedecay
Website: eternal-decay.xyz
Windows: 558889183097d9a991cb2c71b7da3c51
Mac: a4786af0c4ffc84ff193ff2ecbb564b8

Dexis
X: @DexisApp
Website: dexis.app
Notes: Same branding as Swox.

NexVoo
X: @Nexvoospace
Website: nexvoo.app, Nexvoo.net, Nexvoo.us

NexLoop
X: @nexloopspace
Website: nexloop.me

NexoraCore
Notes: Rename of the Nexloop Git repo.

YondaAI
X: @yondaspace
Website: yonda.us

Traffer Groups

A “traffer” malware group is an organized cybercriminal operation that specializes in directing internet users to malicious content typically information-stealing malware through compromised or deceptive websites, ads, and links. They tend to operate in teams with hierarchical structures with administrators recruiting “traffers” (or affiliates) to generate traffic and malware installs via search engine optimization (SEO), YouTube ads, fake software downloads, or owned sites, then monetize the stolen credentials and data via dedicated marketplaces.

A prominent traffer group “CrazyEvil” was identified by Recorded Future in early 2025. The group, who have been active since at least 2021, specialize in social engineering attacks targeted towards cryptocurrency users, influencers, DeFi professionals, and gaming communities. As reported by Recorded Future, CrazyEvil are estimated to have made millions of dollars in revenue from their malicious activity. CrazyEvil and their sub teams create fake software companies, similar to the ones described in this blog, making use of Twitter and Medium to target victims. As seen in this campaign, CrazyEvil instructs users to download their software which is an info stealer targeting both macOS and Windows users.

While it is unclear if the campaigns described in this blog can be attributed to CrazyEvil or any sub teams, the techniques described are similar in nature. This campaign highlights the efforts that threat actors will go to make these fake companies look legitimate in order to steal cryptocurrency from victims, in addition to use of newer evasive versions of malware.

Indicators of Compromise (IoCs)

Manboon[.]com

https://gaetanorealty[.]com

Troveur[.]com

Bigpinellas[.]com

Dsandbox[.]com

Conceptwo[.]com

Aceartist[.]com

turismoelcasco[.]com

Ekodirect[.]com

https://mrajhhosdoahjsd[.]com

https://isnimitz.com/zxc/app[.]zip

http://45[.]94[.]47[.]112/contact

45[.]94[.]47[.]167/contact

77[.]73[.]129[.]18:80

Domain Keys associated with the C2s

YARA Rules

rule Suspicious_Electron_App_Installer

{

  meta:

      description = "Detects Electron apps collecting HWID, MAC, GPU info and executing remote EXEs/MSIs"

      date = "2025-06-18"

  strings:

      $electron_require = /require\(['"]electron['"]\)/

      $axios_require = /require\(['"]axios['"]\)/

      $exec_use = /exec\(.*?\)/

      $url_token = /app-launcher:\/\/.*token=/

      $getHWID = /(Get-CimInstance Win32_ComputerSystemProduct).UUID/

      $getMAC = /details\.mac && details\.mac !== '00:00:00:00:00:00'/

      $getGPU = /wmic path win32_VideoController get name/

      $getInstallDate = /InstallDate/

      $os_info = /os\.cpus\(\)\[0\]\.model/

      $downloadExe = /\.exe['"]/

      $runExe = /msiexec \/i.*\/quiet \/norestart/

      $zipExtraction = /AdmZip\(.*\.extractAllTo/

  condition:

      (all of ($electron_require, $axios_require, $exec_use) and

       3 of ($getHWID, $getMAC, $getGPU, $getInstallDate, $os_info) and

       2 of ($downloadExe, $runExe, $zipExtraction, $url_token))

}

Continue reading
About the author
Tara Gould
Threat Researcher

Blog

/

Identity

/

July 9, 2025

Defending the Cloud: Stopping Cyber Threats in Azure and AWS with Darktrace

fingerprintDefault blog imageDefault blog image

Real-world intrusions across Azure and AWS

As organizations pursue greater scalability and flexibility, cloud platforms like Microsoft Azure and Amazon Web Services (AWS) have become essential for enabling remote operations and digitalizing corporate environments. However, this shift introduces a new set of security risks, including expanding attack surfaces, misconfigurations, and compromised credentials frequently exploited by threat actors.

This blog dives into three instances of compromise within a Darktrace customer’s Azure and AWS environment which Darktrace.

  1. The first incident took place in early 2024 and involved an attacker compromising a legitimate user account to gain unauthorized access to a customer’s Azure environment.
  2. The other two incidents, taking place in February and March 2025, targeted AWS environments. In these cases, threat actors exfiltrated corporate data, and in one instance, was able to detonate ransomware in a customer’s environment.

Case 1 - Microsoft Azure

Simplified timeline of the attack on a customer’s Azure environment.
Figure 1: Simplified timeline of the attack on a customer’s Azure environment.

In early 2024, Darktrace identified a cloud compromise on the Azure cloud environment of a customer in the Europe, the Middle East and Africa (EMEA) region.

Initial access

In this case, a threat actor gained access to the customer’s cloud environment after stealing access tokens and creating a rogue virtual machine (VM). The malicious actor was found to have stolen access tokens belonging to a third-party external consultant’s account after downloading cracked software.

With these stolen tokens, the attacker was able to authenticate to the customer’s Azure environment and successfully modified a security rule to allow inbound SSH traffic from a specific IP range (i.e., securityRules/AllowCidrBlockSSHInbound). This was likely performed to ensure persistent access to internal cloud resources.

Detection and investigation of the threat

Darktrace / IDENTITY recognized that this activity was highly unusual, triggering the “Repeated Unusual SaaS Resource Creation” alert.

Cyber AI Analyst launched an autonomous investigation into additional suspicious cloud activities occurring around the same time from the same unusual location, correlating the individual events into a broader account hijack incident.

Cyber AI Analyst’s investigation into unusual cloud activity performed by the compromised account.
Figure 2: Cyber AI Analyst’s investigation into unusual cloud activity performed by the compromised account.
Figure 2: Surrounding resource creation events highlighted by Cyber AI Analyst.
Figure 3: Surrounding resource creation events highlighted by Cyber AI Analyst.
Figure 4: Surrounding resource creation events highlighted by Cyber AI Analyst.

“Create resource service limit” events typically indicate the creation or modification of service limits (i.e., quotas) for a specific Azure resource type within a region. Meanwhile, “Registers the Capacity Resource Provider” events refer to the registration of the Microsoft Capacity resource provider within an Azure subscription, responsible for managing capacity-related resources, particularly those related to reservations and service limits. These events suggest that the threat actor was looking to create new cloud resources within the environment.

Around ten minutes later, Darktrace detected the threat actor creating or modifying an Azure disk associated with a virtual machine (VM), suggesting an attempt to create a rogue VM within the environment.

Threat actors can leverage such rogue VMs to hijack computing resources (e.g., by running cryptomining malware), maintain persistent access, move laterally within the cloud environment, communicate with command-and-control (C2) infrastructure, and stealthily deliver and deploy malware.

Persistence

Several weeks later, the compromised account was observed sending an invitation to collaborate to an external free mail (Google Mail) address.

Darktrace deemed this activity as highly anomalous, triggering a compliance alert for the customer to review and investigate further.

The next day, the threat actor further registered new multi-factor authentication (MFA) information. These actions were likely intended to maintain access to the compromised user account. The customer later confirmed this activity by reviewing the corresponding event logs within Darktrace.

Case 2 – Amazon Web Services

Simplified timeline of the attack on a customer’s AWS environment
Figure 5: Simplified timeline of the attack on a customer’s AWS environment

In February 2025, another cloud-based compromised was observed on a UK-based customer subscribed to Darktrace’s Managed Detection and Response (MDR) service.

How the attacker gained access

The threat actor was observed leveraging likely previously compromised credential to access several AWS instances within customer’s Private Cloud environment and collecting and exfiltrating data, likely with the intention of deploying ransomware and holding the data for ransom.

Darktrace alerting to malicious activity

This observed activity triggered a number of alerts in Darktrace, including several high-priority Enhanced Monitoring alerts, which were promptly investigated by Darktrace’s Security Operations Centre (SOC) and raised to the customer’s security team.

The earliest signs of attack observed by Darktrace involved the use of two likely compromised credentials to connect to the customer’s Virtual Private Network (VPN) environment.

Internal reconnaissance

Once inside, the threat actor performed internal reconnaissance activities and staged the Rclone tool “ProgramData\rclone-v1.69.0-windows-amd64.zip”, a command-line program to sync files and directories to and from different cloud storage providers, to an AWS instance whose hostname is associated with a public key infrastructure (PKI) service.

The threat actor was further observed accessing and downloading multiple files hosted on an AWS file server instance, notably finance and investment-related files. This likely represented data gathering prior to exfiltration.

Shortly after, the PKI-related EC2 instance started making SSH connections with the Rclone SSH client “SSH-2.0-rclone/v1.69.0” to a RockHoster Virtual Private Server (VPS) endpoint (193.242.184[.]178), suggesting the threat actor was exfiltrating the gathered data using the Rclone utility they had previously installed. The PKI instance continued to make repeated SSH connections attempts to transfer data to this external destination.

Darktrace’s Autonomous Response

In response to this activity, Darktrace’s Autonomous Response capability intervened, blocking unusual external connectivity to the C2 server via SSH, effectively stopping the exfiltration of data.

This activity was further investigated by Darktrace’s SOC analysts as part of the MDR service. The team elected to extend the autonomously applied actions to ensure the compromise remained contained until the customer could fully remediate the incident.

Continued reconissance

Around the same time, the threat actor continued to conduct network scans using the Nmap tool, operating from both a separate AWS domain controller instance and a newly joined device on the network. These actions were accompanied by further internal data gathering activities, with around 5 GB of data downloaded from an AWS file server.

The two devices involved in reconnaissance activities were investigated and actioned by Darktrace SOC analysts after additional Enhanced Monitoring alerts had triggered.

Lateral movement attempts via RDP connections

Unusual internal RDP connections to a likely AWS printer instance indicated that the threat actor was looking to strengthen their foothold within the environment and/or attempting to pivot to other devices, likely in response to being hindered by Autonomous Response actions.

This triggered multiple scanning, internal data transfer and unusual RDP alerts in Darktrace, as well as additional Autonomous Response actions to block the suspicious activity.

Suspicious outbound SSH communication to known threat infrastructure

Darktrace subsequently observed the AWS printer instance initiating SSH communication with a rare external endpoint associated with the web hosting and VPS provider Host Department (67.217.57[.]252), suggesting that the threat actor was attempting to exfiltrate data to an alternative endpoint after connections to the original destination had been blocked.

Further investigation using open-source intelligence (OSINT) revealed that this IP address had previously been observed in connection with SSH-based data exfiltration activity during an Akira ransomware intrusion [1].

Once again, connections to this IP were blocked by Darktrace’s Autonomous Response and subsequently these blocks were extended by Darktrace’s SOC team.

The above behavior generated multiple Enhanced Monitoring alerts that were investigated by Darktrace SOC analysts as part of the Managed Threat Detection service.

Enhanced Monitoring alerts investigated by SOC analysts as part of the Managed Detection and Response service.
Figure 5: Enhanced Monitoring alerts investigated by SOC analysts as part of the Managed Detection and Response service.

Final containment and collaborative response

Upon investigating the unusual scanning activity, outbound SSH connections, and internal data transfers, Darktrace analysts extended the Autonomous Response actions previously triggered on the compromised devices.

As the threat actor was leveraging these systems for data exfiltration, all outgoing traffic from the affected devices was blocked for an additional 24 hours to provide the customer’s security team with time to investigate and remediate the compromise.

Additional investigative support was provided by Darktrace analysts through the Security Operations Service, after the customer's opened of a ticket related to the unfolding incident.

Simplified timeline of the attack
Figure 8: Simplified timeline of the attack

Around the same time of the compromise in Case 2, Darktrace observed a similar incident on the cloud environment of a different customer.

Initial access

On this occasion, the threat actor appeared to have gained entry into the AWS-based Virtual Private Cloud (VPC) network via a SonicWall SMA 500v EC2 instance allowing inbound traffic on any port.

The instance received HTTPS connections from three rare Vultr VPS endpoints (i.e., 45.32.205[.]52, 207.246.74[.]166, 45.32.90[.]176).

Lateral movement and exfiltration

Around the same time, the EC2 instance started scanning the environment and attempted to pivot to other internal systems via RDP, notably a DC EC2 instance, which also started scanning the network, and another EC2 instance.  

The latter then proceeded to transfer more than 230 GB of data to the rare external GTHost VPS endpoint 23.150.248[.]189, while downloading hundreds of GBs of data over SMB from another EC2 instance.

Cyber AI Analyst incident generated following the unusual scanning and RDP connections from the initial compromised device.
Figure 7: Cyber AI Analyst incident generated following the unusual scanning and RDP connections from the initial compromised device.

The same behavior was replicated across multiple EC2 instances, whereby compromised instances uploaded data over internal RDP connections to other instances, which then started transferring data to the same GTHost VPS endpoint over port 5000, which is typically used for Universal Plug and Play (UPnP).

What Darktrace detected

Darktrace observed the threat actor uploading a total of 718 GB to the external endpoint, after which they detonated ransomware within the compromised VPC networks.

This activity generated nine Enhanced Monitoring alerts in Darktrace, focusing on the scanning and external data activity, with the earliest of those alerts triggering around one hour after the initial intrusion.

Darktrace’s Autonomous Response capability was not configured to act on these devices. Therefore, the malicious activity was not autonomously blocked and escalated to the point of ransomware detonation.

Conclusion

This blog examined three real-world compromises in customer cloud environments each illustrating different stages in the attack lifecycle.

The first case showcased a notable progression from a SaaS compromise to a full cloud intrusion, emphasizing the critical role of anomaly detection when legitimate credentials are abused.

The latter two incidents demonstrated that while early detection is vital, the ability to autonomously block malicious activity at machine speed is often the most effective way to contain threats before they escalate.

Together, these incidents underscore the need for continuous visibility, behavioral analysis, and machine-speed intervention across hybrid environments. Darktrace's AI-driven detection and Autonomous Response capabilities, combined with expert oversight from its Security Operations Center, give defenders the speed and clarity they need to contain threats and reduce operational disruption, before the situation spirals.

Credit to Alexandra Sentenac (Senior Cyber Analyst) and Dylan Evans (Security Research Lead)

References

[1] https://www.virustotal.com/gui/ip-address/67.217.57.252/community

Case 1

Darktrace / IDENTITY model alerts

IaaS / Compliance / Uncommon Azure External User Invite

SaaS / Resource / Repeated Unusual SaaS Resource Creation

IaaS / Compute / Azure Compute Resource Update

Cyber AI Analyst incidents

Possible Unsecured AzureActiveDirectory Resource

Possible Hijack of Office365 Account

Case 2

Darktrace / NETWORK model alerts

Compromise / SSH Beacon

Device / Multiple Lateral Movement Model Alerts

Device / Suspicious SMB Scanning Activity

Device / SMB Lateral Movement

Compliance / SSH to Rare External Destination

Device / Anomalous SMB Followed By Multiple Model Alerts

Device / Anonymous NTLM Logins

Anomalous Connection / SMB Enumeration

Device / New or Uncommon SMB Named Pipe Device / Network Scan

Device / Suspicious Network Scan Activity

Device / New Device with Attack Tools

Device / RDP Scan Device / Attack and Recon Tools

Compliance / High Priority Compliance Model Alert

Compliance / Outgoing NTLM Request from DC

Compromise / Large Number of Suspicious Successful Connections

Device / Large Number of Model Alerts

Anomalous Connection / Multiple Failed Connections to Rare Endpoint

Unusual Activity / Internal Data Transfer

Anomalous Connection / Unusual Internal Connections

Device / Anomalous RDP Followed By Multiple Model Alerts

Unusual Activity / Unusual External Activity

Unusual Activity / Enhanced Unusual External Data Transfer

Unusual Activity / Unusual External Data Transfer

Unusual Activity / Unusual External Data to New Endpoint

Anomalous Connection / Multiple Connections to New External TCP Port

Darktrace / Autonomous Response model alerts

Antigena / Network / Significant Anomaly / Antigena Enhanced Monitoring from Server Block

Antigena / Network / Manual / Quarantine Device

Antigena / MDR / MDR-Quarantined Device

Antigena / MDR / Model Alert on MDR-Actioned Device

Antigena / Network / Significant Anomaly / Antigena Enhanced Monitoring from Client Block

Antigena / Network / Significant Anomaly / Antigena Alerts Over Time Block

Antigena / Network / Insider Threat / Antigena Network Scan Block

Antigena / Network / Significant Anomaly / Antigena Significant Server Anomaly Block

Antigena / Network / Insider Threat / Antigena SMB Enumeration Block

Antigena / Network / Significant Anomaly / Antigena Controlled and Model Alert

Antigena / Network / Significant Anomaly / Antigena Significant Anomaly from Client Block

Antigena / Network / External Threat / Antigena Suspicious Activity Block

Antigena / Network / Insider Threat / Antigena Internal Data Transfer Block

Cyber AI Analyst incidents

Possible Application Layer Reconnaissance Activity

Scanning of Multiple Devices

Unusual Repeated Connections

Unusual External Data Transfer

Case 3

Darktrace / NETWORK model alerts

Unusual Activity / Unusual Large Internal Transfer

Compliance / Incoming Remote Desktop

Unusual Activity / High Volume Server Data Transfer

Unusual Activity / Internal Data Transfer

Anomalous Connection / Unusual Internal Remote Desktop

Anomalous Connection / Unusual Incoming Data Volume

Anomalous Server Activity / Domain Controller Initiated to Client

Device / Large Number of Model Alerts

Anomalous Connection / Possible Flow Device Brute Force

Device / RDP Scan

Device / Suspicious Network Scan Activity

Device / Network Scan

Anomalous Server Activity / Anomalous External Activity from Critical Network Device

Anomalous Connection / Download and Upload

Unusual Activity / Unusual External Data Transfer

Unusual Activity / High Volume Client Data Transfer

Unusual Activity / Unusual External Activity

Anomalous Connection / Uncommon 1 GiB Outbound

Device / Increased External Connectivity

Compromise / Large Number of Suspicious Successful Connections

Anomalous Connection / Data Sent to Rare Domain

Anomalous Connection / Low and Slow Exfiltration to IP

Unusual Activity / Enhanced Unusual External Data Transfer

Anomalous Connection / Multiple Connections to New External TCP Port

Anomalous Server Activity / Outgoing from Server

Anomalous Connection / Multiple Connections to New External UDP Port

Anomalous Connection / Possible Data Staging and External Upload

Unusual Activity / Unusual External Data to New Endpoint

Device / Large Number of Model Alerts from Critical Network Device

Compliance / External Windows Communications

Anomalous Connection / Unusual Internal Connections

Cyber AI Analyst incidents

Scanning of Multiple Devices

Extensive Unusual RDP Connections

MITRE ATT&CK mapping

(Technique name – Tactic ID)

Case 1

Defense Evasion - Modify Cloud Compute Infrastructure: Create Cloud Instance

Persistence – Account Manipulation

Case 2

Initial Access - External Remote Services

Execution - Inter-Process Communication

Persistence - External Remote Services

Discovery - System Network Connections Discovery

Discovery - Network Service Discovery

Discovery - Network Share Discovery

Lateral Movement - Remote Desktop Protocol

Lateral Movement - Remote Services: SMB/Windows Admin Shares

Collection - Data from Network Shared Drive

Command and Control - Protocol Tunneling

Exfiltration - Exfiltration Over Asymmetric Encrypted Non-C2 Protocol

Case 3

Initial Access - Exploit Public-Facing Application

Discovery - Remote System Discovery

Discovery - Network Service Discovery

Lateral Movement - Remote Services

Lateral Movement - Remote Desktop Protocol  

Collection - Data from Network Shared Drive

Collection - Data Staged: Remote Data Staging

Exfiltration - Exfiltration Over C2 Channel

Command and Control - Non-Standard Port

Command and Control – Web Service

Impact - Data Encrypted for Impact

List of IoCs

IoC         Type      Description + Probability

193.242.184[.]178 - IP Address - Possible Exfiltration Server  

45.32.205[.]52  - IP Address  - Possible C2 Infrastructure

45.32.90[.]176 - IP Address - Possible C2 Infrastructure

207.246.74[.]166 - IP Address - Likely C2 Infrastructure

67.217.57[.]252 - IP Address - Likely C2 Infrastructure

23.150.248[.]189 - IP Address - Possible Exfiltration Server

Continue reading
About the author
Alexandra Sentenac
Cyber Analyst
Your data. Our AI.
Elevate your network security with Darktrace AI