Introduction

Among the most significant events in the AI world in early 2025 was the release of DeepSeek-R1 – a powerful reasoning large language model (LLM) with open weights. It’s available both for local use and as a free service. Since DeepSeek was the first service to offer access to a reasoning LLM to a wide audience, it quickly gained popularity, mirroring the success of ChatGPT. Naturally, this surge in interest also attracted cybercriminals.

While analyzing our internal threat intelligence data, we discovered several groups of websites mimicking the official DeepSeek chatbot site and distributing malicious code disguised as a client for the popular service.

Screenshot of the official DeepSeek website (February 2025)

Scheme 1: Python stealer and non-existent DeepSeek client

The first group of websites was hosted on domains whose names included DeepSeek model versions (V3 and R1):

• r1-deepseek[.]net;
• v3-deepseek[.]com.

As shown in the screenshot, the fake website lacks the option to start a chat – you can only download an application. However, the real DeepSeek doesn’t have an official Windows client.

Screenshot of the fake website

Clicking the “Get DeepSeek App” button downloads a small archive,
deep-seek-installation.zip. The archive contains the DeepSeek Installation.lnk file, which holds a URL.
At the time of publishing this research, the attackers had modified the fake page hosted on the
v3-deepseek[.]com domain. It now prompts users to download a client for the Grok model developed by xAI. We’re observing similar activity on the v3-grok[.]com domain as well. Disguised as a client is an archive named grok-ai-installation.zip, containing the same shortcut.
Executing the .lnk file runs a script located at the URL inside the shortcut:

This script downloads and unpacks an archive named
f.zip.

Contents of the unpacked archive

Next, the script runs the
1.bat file from the unpacked archive.

Contents of the BAT file

The downloaded archive also contains the
svchost.exe and python.py files. The first one is a legitimate file python.exe, renamed to mimic a Windows process to mislead users checking running applications in Task Manager.

It is used to launch
python.py, which contains the malicious payload (we’ve also seen this file named code.py). This is a stealer script written in Python that we haven’t seen in attacks before. If it’s executed successfully, the attackers obtain a wealth of data from the victim’s computer: cookies and session tokens from various browsers, login credentials for email, gaming, and other accounts, files with certain extensions, cryptocurrency wallet information, and more.

After collecting the necessary data, the script generates an archive and then either sends it to the stealer’s operators using a Telegram bot or uploads it to the Gofile file-sharing service. Thus, attempting to use the chatbot could result in the victim losing social media access, personal data, and even cryptocurrency. If corporate credentials are stored on the compromised device, entire organizations could also be at risk, leading to far more severe consequences.

Scheme 2: Malicious script and a million views

In another case, fake DeepSeek websites were found on the following domains:

• deepseek-pc-ai[.]com
• deepseek-ai-soft[.]com

We discovered the first domain back in early February, hosting the default Apache web server page with no content. Later, this domain displayed a new web page closely resembling the DeepSeek website. Notably, the fake site uses geofencing: when requests come from certain IP addresses, such as Russian ones, it returns a placeholder page filled with generic SEO text about DeepSeek (we believe this text may have been LLM-generated):

If the IP address and other request parameters meet the specified criteria, the server returns a page resembling DeepSeek. Users are prompted to download a client or start the chatbot, but either action results in downloading a malicious installer created using Inno Setup. Kaspersky products detect it as
Trojan-Downloader.Win32.TookPS.*.

When executed, this installer contacts malicious URLs to receive a command that will be executed using cmd. The most common command launches
powershell.exe with a Base64-encoded script as an argument. This script accesses an encoded URL to download another PowerShell script, which activates the built-in SSH service and modifies its configuration using the attacker’s keys, allowing remote access to the victim’s computer.

Part of the malicious PowerShell script

This case is notable because we managed to identify the primary vector for spreading the malicious links – posts on the social network X (formerly Twitter):

This post, directing users to
deepseek-pc-ai[.]com, was made from an account belonging to an Australian company. The post gained 1.2 million views and over a hundred reposts, most of which were probably made by bots – note the similar usernames and identifiers in their bios:

Some users in the comments dutifully point out the malicious nature of the link.

Links to
deepseek-ai-soft[.]com were also distributed through X posts, but at the time of investigation, they were only available in Google’s cache:

Scheme 3: Backdoors and attacks on Chinese users

We also encountered sites that directly distributed malicious executable files. One such file was associated with the following domains:

• app.delpaseek[.]com;
• app.deapseek[.]com;
• dpsk.dghjwd[.]cn.

These attacks target more technically advanced users – the downloaded malicious payload mimics Ollama, a framework for running LLMs such as DeepSeek on local hardware. This tactic reduces suspicion among potential victims. Kaspersky solutions detect this payload as
Backdoor.Win32.Xkcp.a.

The victim only needed to launch the “DeepSeek client” on their device to trigger the malware, which creates a KCP tunnel with predefined parameters.

Additionally, we observed attacks where a victim’s device downloaded the
deep_windows_Setup.zip archive, containing a malicious executable. The archive was downloaded from the following domains:

• deep-seek[.]bar;
• deep-seek[.]rest.

The malware in the archive is detected by Kaspersky solutions as
Trojan.Win32.Agent.xbwfho. This is an installer created with Inno Setup that uses DLL sideloading to load a malicious library. The DLL in turn extracts and loads into memory a payload hidden using steganography — a Farfli backdoor modification — and injects it into a process.
Both of these campaigns, judging by the language of the bait pages, are targeting Chinese-speaking users.

Conclusion

The nature of the fake websites described in this article suggests these campaigns are widespread and not aimed at specific users.

Cybercriminals use various schemes to lure victims to malicious resources. Typically, links to such sites are distributed through messengers and social networks, as seen in the example with the X post. Attackers may also use typosquatting or purchase ad traffic to malicious sites through numerous affiliate programs.

We strongly advise users to carefully check the addresses of websites they visit, especially if links come from unverified sources. This is especially important for highly popular services. In this case, it’s particularly noteworthy that DeepSeek doesn’t have a native Windows client. This isn’t the first time that cybercriminals have exploited the popularity of chatbots to distribute malware: they’ve previously targeted regular users with Trojans disguised as ChatGPT clients and developers with malicious packages in PyPI. Simple digital hygiene practices, combined with a cutting-edge security solution, can significantly reduce the risk of device infection and personal data loss.

Indicators of compromise

MD5

4ef18b2748a8f499ed99e986b4087518
155bdb53d0bf520e3ae9b47f35212f16
6d097e9ef389bbe62365a3ce3cbaf62d
3e5c2097ffb0cb3a6901e731cdf7223b
e1ea1b600f218c265d09e7240b7ea819
7cb0ca44516968735e40f4fac8c615ce
7088986a8d8fa3ed3d3ddb1f5759ec5d

Malicious domains

r1-deepseek[.]net
v3-deepseek[.]com
deepseek-pc-ai[.]com
deepseek-ai-soft[.]com
app.delpaseek[.]com
app.deapseek[.]com
dpsk.dghjwd[.]cn
deep-seek[.]bar
deep-seek[.]rest
v3-grok[.]com

securelist.com/backdoors-and-s…

For those who love systems and structure, owning a 19-inch rack with just one slot filled is just not it. But what if the rest of your gear isn’t 19-inch? Well, then you go out and make it so, just like [Cal Bryant] did recently.

The goal was to consolidate multiple devices — DAC, input selector, streamer, and power routing — into a single 2U rackmount unit. His first attempts involved drilling 1U panels to attach gear with removable faceplates. That worked, but not all devices played nice. So his next step became a fully custom enclosure with CAD-modeled brackets and front panels.

OpenSCAD turned out to be a lifesaver, letting [Cal] design modular mounting solutions. Exporting proper circles for CNC turret punching however appeared to be a nightmare. It was FreeCAD to the rescue for post-processing. After some sanding and auto-shop painting, the final faceplate looked factory-made.

Custom switch boxes for power and audio routing keep things tidy, housing everything from USB to XLR inputs. A 4-pole switch even allows seamless swapping between his DAC and DJ controller, while UV-printed graphics bring the finishing touch to this project. For those looking to clean up their Hi-Fi setup (or just love modding for the sake of it), there’s a lot to learn from this build.

If buying a rack is not within your budget, you could start with well-known IKEA LACK furniture.

hackaday.com/2025/03/06/rackmo…

peertube.uno/w/g6JCBVHtnit3aZe…

Maurizio Delladio

2025-03-05 19:49:41

Parola Di Berlinguer

Ieri, 4 marzo 2025, presso il Teatro Portaportese in via Portuense a Roma, si è tenuta una iniziativa curata da Sergio Gentili e Lelio La Porta (voce narrante). La stessa si è articolata in una parte musicale, in alcuni ricordi/narrazioni legate agli anni tra il 1969 e 1984 ed in letture di articoli ed interviste di Enrico Berlinguer

Vorrei condividere con voi tre canzoni che porto nel cuore da più o meno tempo ma che ritengo assai interessanti e che forse nell'italia di oggi mancano di essere conosciute perché non molto condivise sui social ma abbastanza conosciute nel resto del mondo e apprezzate davveeo tanto. Ritengo che mancano ancora tanto artisti di questo livello in italia e che spesso manca l'ascolto dalle persone distratte da tutto questo intrattenimento e ci vorrebbe una pausa per l'ascolto, pausa che è sostanzialmente la stessa concezione che danno anche i buddisti/mindfuless ed affini quando si dovrebbe parlare di vita, prezenza, di ascolto di se stessi e del corpo/cuore.

Ho parlato/scritto troppo, vi lascio alle tre canzoni da ca(r)pire da soli, vado via! *sciò*

- Delilah Montagu - Version Of Me -

(canzone riflessiva su chi si è veramente e cosa si comunica davvero se stessi della stessa artista di cui fred again ha campionato la sua voce per un suo pezzo molto famoso)

youtube.com/watch?v=wema_tYrK3…

- Regina Spektor - Loveology -

(consiglio l'ascolto in cuffia con bassissime aspettative arrivando al finale che vi stupirà di sicuro...non demordete che è un pezzo davvero incredibile che ho in loop da ore)

youtube.com/watch?v=lHaetZ-Vu8…

- Raphael Treza - We Are One -

(soundtrack favolosa con un testo poetico fantastico di un documentario di raphael treza trovabile facilmente free su youtube)
(raphaeltreza.com/)

youtube.com/watch?v=kSSlt7Uf_K…

C'era una volta un filo d'argento, sottile e lucente, che giaceva silenzioso in un angolo del laboratorio. Non era un filo come tanti, era un filo speciale, scelto con cura per la sua purezza e la sua capacità di catturare la luce. Un filo che aspettava di prendere vita, di trasformarsi in qualcosa di unico e prezioso.

Specifiche e curiosità 👇

solaraartigianatoartistico.wor…