Although direct messages sent through a chat app are often perceived as a private conversation, it’s actually not that simple. Not only can your chats and data be used for advertising and AI training, but they can be shared with law enforcement and intelligence agencies. Furthermore, perfect strangers, or scammers — pretending to be your boss, for example — might reach out to you directly. Then again, attackers can use social engineering techniques to gain access to your account and read all of your chats in real-time.

Which services minimize the chance of these unwelcome events? That’s the question experts at the company Incogni set out to answer. They decided to compare popular social networks and messaging apps, and ranked them as per privacy levels from highest to lowest. The result is the Social Media Privacy Ranking 2025. This was an extensive study covering 15 social networks and chat apps, and comparing them across 18 criteria. Today, we focus on the scores for messaging apps and direct communication platforms — selecting only the most practical evaluation criteria. So, which of the common messaging apps are the most privacy-oriented?

Overall privacy rankings

We’ll start with Incogni’s final conclusions. After summing up all their scores across the criteria, they produced the following privacy rankings (lower is better):

1. Discord: 10.23
2. Telegram: 13.08
3. Snapchat: 13.39
4. Facebook Messenger: 22.22
5. WhatsApp: 23.17

But don’t rush to migrate all your chats from WhatsApp to Discord just yet — comparing only the criteria that matter most reveals a different picture. Incogni’s comprehensive study included some very peculiar points, such as the number of fines for data retention violations across all countries, the number of past hacks and data breaches, the readability of the privacy policy, the time it takes to have your data deleted after an account closure request, and so on.

However, there are also highly practical criteria: the types of data collected by the mobile app, the privacy level by default, the amount of user data visible to non-contacts, the use of user data for AI training, and the option to opt out of this. For those concerned about excessive government interference in private correspondence, the score for the response rates to government requests for user information will also be of interest. If we add up the scores from only these practical categories, the rankings shift significantly:

1. Telegram: 4.23
2. Snapchat: 7.72
3. Discord: 8.14
4. WhatsApp: 11.93
5. Facebook Messenger: 13.37

Incogni penalized WhatsApp  3.4 points for the fact that chats may be used for AI training, and users can’t opt out. However, there’s one important caveat: as of today, this only applies to user chats with Meta’s AI assistant, while other chats are still protected by end-to-end encryption, and can’t be used for training. Therefore, in our view, a more accurate score for WhatsApp would be 8.53; this doesn’t change its position, but significantly narrows the gap between it and the leading trio.

Let’s move past the numbers now, and review the practically significant findings of the analysis.

Private by default

An app focused on user interests sets all security and privacy settings to safe and private upon installation. The user can then lower the level of privacy where they choose to. Telegram and Snapchat exhibit this commendable behavior. Discord’s default settings are less private, while Facebook Messenger and WhatsApp are down at the bottom of the rankings. A similar situation is found with the number of privacy settings — Telegram and Snapchat offer the most.

We’ve published detailed guides on setting up privacy in Telegram, WhatsApp, and Discord, and you can find privacy configuration tips for many other popular apps, devices, and operating systems on our free Privacy Checker portal.

Secure against strangers

Minimizing the amount of information strangers can see is crucial for both privacy and physical safety. It limits the possibilities for scams, spam, stalking, and child abuse. The most secure accounts are provided equally by Telegram and WhatsApp — tying for first place. Facebook Messenger and Snapchat share second, while Discord ranks last in this regard.

Cooperating with authorities

Telegram doesn’t disclose the percentage of government requests for personal data that it grants — though it’s known to be greater than zero. As for the other platforms, Snapchat most frequently approves such requests (82%), Meta’s services approve them in 78% of cases (the breakdown by service is unknown), and Discord is not far behind at 77.4%.

Collecting data for advertising and other purposes

Every platform collects a certain amount of information about its users, their socio-demographic profile, and preferences. The study distinguishes between general data collection and mobile-app data collection. The former was based on privacy policies; the latter used the data published for the apps in the App Store and Google Play.

Based on general data collection, the leaders with the least amount of collected data are Telegram and WhatsApp. Discord took second place, and Snapchat and Facebook Messenger both ranked last. Regarding mobile-app data collection, the picture is slightly different: Telegram leads by a significant margin, followed by WhatsApp in second place, then Discord, Snapchat, and finally, Facebook Messenger.

Which messaging app is best?

Among the services reviewed, Telegram collects the least data and provides the widest range of privacy settings. While Discord leads the overall rankings thanks to limited data collection and a clean record on privacy fines, it falls short in privacy settings, and doesn’t default to secure options. WhatsApp offers extensive protection against strangers, and collects a relatively modest amount of user data.

Note that the ranking focuses on mainstream apps; more niche messaging apps that place a strong emphasis on privacy were simply not included. Truly confidential/sensitive conversations should ideally be conducted on one of these dedicated private messaging apps.

Additionally, Incogni didn’t focus on encryption. Among the reviewed apps, only WhatsApp offers full end-to-end encryption for all chats by default. This is a crucial consideration, for the hugely popular Telegram doesn’t guarantee message privacy: chats aren’t end-to-end encrypted by default.

Finally, don’t forget that the indicated level of security applies only to the official mobile clients of these messaging services. The desktop versions of popular messaging apps are far more vulnerable due to their architecture. As for using mods or third-party clients, it’s best to avoid them entirely — malicious versions are routinely distributed both through channels and group chats within the messaging services themselves, and through official app stores such as Google Play.

To protect Android smartphones from these malicious apps, consider Kaspersky for Android. Incidentally, after a recent update, it now also blocks phishing and malicious links in all notifications from any messaging or other app.

Messaging apps today arguably hold the maximum amount of private information about each of us. To avoid becoming a victim of a data leak, read our other posts:

• Messengers 101: safety and privacy advice
• What to do if your WhatsApp account gets hacked
• What to do if your Telegram account is hacked
• WhatsApp and Telegram account hijacking: how to protect yourself against scams
• Chatting offline: an overview of mesh messaging apps

Kaspersky official blog – ​Read More

Here at ANY.RUN, we know how crucial threat intelligence is for ensuring strong cybersecurity, especially in organizations. 

This year, our efforts in promoting this data-driven approach to solving the needs of businesses were praised at CyberSecurity Breakthrough Awards. ANY.RUN was recognized as the Threat Intelligence Company of the Year 2025. 

New Milestone on the Way to Safer Future 

This wasn’t an easy win: the CyberSecurity Breakthrough Awards is a prestigious international program with an independent panel of industry experts in the jury. Our sincere thanks go to them for acknowledging our impact on leading innovative enterprise-grade solutions forward. 

But above all, we’d like to thank our global community of clients, contributors, and partners. It’s a shared win for all of us. 

Threat Intelligence Changes Everything 

Earlier this year, ANY.RUN’s solutions gained global acclaim and won multiple awards, like Globee Awards and Cybersecurity Excellence Awards. But this victory stands out, as it recognizes our influence as a company and reflects our approach focused on the integrity of a unified workflow.  

ANY.RUN’s Threat Intelligence Feeds and Threat Intelligence Lookup are redefining how SOCs operate in today’s threat landscape. Instead of relying on outdated indicators from post-incident reports, ANY.RUN leverages insights from a global community of over 500,000 analysts and 15,000 organizations actively analyzing the latest threats in our Interactive Sandbox. 

 Continuously updated threat intelligence helps SOC teams gain automation and data needed to stay ahead of evolving attacks. Security leaders can make faster, more confident decisions, strengthen proactive defense strategies, and maximize the ROI of their security stack. 

• Access 24x more IOCs per incident for wider visibility: Live data on global attacks ensures comprehensive threat coverage of new malware and phishing. 

• Enrich your system with 99% unique IOCs to reduce workload: In-depth intel cuts Tier 1/Tier 2 investigations and promotes confident decisions. 

• Accelerate MTTR by 21 min per case for faster action: Threat behavior context for IOCs/IOAs/IOBs provides insights for streamlined incident mitigation. 

About ANY.RUN 

Over 500,000 cybersecurity professionals and 15,000+ organizations across finance, manufacturing, healthcare, and other industries trust ANY.RUN and accelerate their malware investigations worldwide. 

Faster triage and response with ANY.RUN’s Interactive Sandbox: Safely detonate suspicious files, observe malicious behavior in real time, and gain insights for faster, confident security decisions. 

Enhance threat detection with Threat Intelligence Lookup and Threat Intelligence Feeds: Tap into actionable, up-to-date intelligence to detect and understand emerging threats instantly. 

Experience how ANY.RUN’s solutions can power your SOC: start 14-day trial 

The post ANY.RUN Recognized as Threat Intelligence Company of the Year 2025   appeared first on ANY.RUN’s Cybersecurity Blog.

ANY.RUN’s Cybersecurity Blog – ​Read More

Welcome to this week’s edition of the Threat Source newsletter. 

“The truth about the world, he said, is that anything is possible… For existence has its own order and that no man’s mind can compass, that mind itself being but a fact among others.” ― Cormac McCarthy, “Blood Meridian”

Earlier this week, I spent a few days off to take a reading retreat with my wife. Diving into several books from various genres and sitting on several quiet acres in the Texas hill country was a wonderful way to refuel. While reading a completely different book, I was reminded of this quoted section from Cormac and it gave me pause.  

We, as security practitioners, often move forward with the knowledge and expertise we’ve acquired along the various paths that led us to this point. It’s easy to fall into the trap of assuming that, because we’ve shared similar experiences, we all possess the same skillsets — each of us following our own string in the maze. In the end, the gaps between what we assume about each other and the reality can be tremendous. How do we ensure these aren’t the very gaps adversaries exploit, and that our perceived strengths don’t become our weaknesses? 

It comes down to communication and community. Talking openly about our skillsets can feel unnecessary among seasoned practitioners—we often assume that everyone has followed a similar path through the maze of their careers. But in practice, conversations quickly reveal this isn’t the case. One of the best ways to build the soft skills that help your career grow is by meeting with your team and discussing the technical skills that got you here and how you apply them now, which skills still benefit your daily routine, and which have fallen to the wayside via obsolescence. If you can create a meeting focused on this topic — rotating among your direct team and involving the team or leader above — you’ll start to pinpoint the skillsets that have the greatest impact on your day-to-day work.  

This process will help you identify specific technical skills needed for hiring or training new team members. It also gives junior analysts a clear view of what senior analysts rely on in their expanded roles, helping to guide their own education. Furthermore, when everyone understands the technical and soft skills that leadership uses, it can help remove the distance between technical and people leaders — and can leave a “string” for those early in their careers to follow as they navigate their own path through the maze 

Know your environment. This is always my first answer whenever I’m asked what to do next or how to proceed. Knowing with extreme clarity both the skillsets on your team and the foundation they’re built upon will allow your team to thrive and grow. It also makes it easier to identify opportunities for cross-training and to provide targeted mentorship where it’s needed most. 

“War was always here. Before man was, war waited for him. The ultimate trade awaiting its ultimate practitioner.” ― Cormac McCarthy, “Blood Meridian”

The one big thing 

According to the Cisco Talos IR Trends Q3 2025 report, over 60% of our incident response cases involved attackers exploiting public-facing applications, mainly through the ToolShell attack chain against unpatched Microsoft SharePoint servers. This is a huge jump from under 10% last quarter. About 20% of cases were ransomware-related (down from 50%), but new ransomware variants and tactics, like using legitimate tools for persistence, were seen. Attackers also increased their use of compromised internal accounts for phishing, and public administration became the top targeted sector for the first time since we began these reports in 2021. 

Why do I care? 

Attackers are going after anyone with exposed or outdated systems, including local governments and public services. With attackers exploiting new vulnerabilities almost immediately (especially in widely used software like SharePoint), even small delays in patching or weak internal defenses can put your organization and its data at serious risk. 

So now what? 

Prioritize rapid patching of public-facing applications, especially after new vulnerabilities are disclosed, and implement strong network segmentation to limit attackers’ lateral movement. Additionally, enhancing multi-factor authentication, improving centralized logging, and educating your users can help detect and block attacks earlier. 

Top security headlines of the week 

Vulnerability in Dolby Decoder can allow zero-click attacks 
Tracked as CVE-2025-54957 (CVSS score of 7.0), the security defect can be triggered using malicious audio messages, leading to remote code execution. On Android, the vulnerability can be exploited remotely without user interaction. (SecurityWeek) 

131 Chrome extensions caught hijacking WhatsApp Web for massive spam campaign 
Researchers have uncovered a coordinated campaign that leveraged 131 rebranded clones of a WhatsApp Web automation extension for Google Chrome to spam Brazilian users at scale. (The Hacker News) 

Prosper data breach impacts 17.6 million accounts 
Hackers stole personal and financial details belonging to 17.6 million users of the Prosper lending platform, including Social Security numbers and government IDs. (SecurityWeek) 

“PassiveNeuron” cyber spies target orgs with custom malware 
A threat campaign is targeting high-profile organizations in the government, industrial, and financial sectors across Asia, Africa, and Latin America, with two custom malware implants designed for cyber espionage. (Dark Reading) 

Can’t get enough Talos? 

• Reducing abuse of Microsoft 365 Exchange Online Direct Send 
  Cisco Talos has observed increased activity by malicious actors leveraging Direct Send as part of phishing campaigns. Here’s how to strengthen your defenses. 
• Beers with Talos: Two Marshalls, one podcast 
  Talos’ Vice President Christopher Marshall (the “real Marshall,” much to Joe’s displeasure) joins Hazel, Bill, and Joe for a very real conversation about leading people when the world won’t stop moving. 

Upcoming events where you can find Talos 

• NTNU MALWAREFORUM (Oct. 28 – 29) Oslo, Norway 
• Bsides Osijek (Nov. 5) Osijek, Croatia 
• AVAR (Dec. 3 – 5) Kuala Lumpur, Malaysia 

Most prevalent malware files from Talos telemetry over the past week 

SHA256: d933ec4aaf7cfe2f459d64ea4af346e69177e150df1cd23aad1904f5fd41f44a  
MD5: 1f7e01a3355b52cbc92c908a61abf643 
Talos Rep: https://talosintelligence.com/talos_file_reputation?s=d933ec4aaf7cfe2f459d64ea4af346e69177e150df1cd23aad1904f5fd41f44a  
Example Filename: cleanup.bat  
Detection Name: W32.D933EC4AAF-90.SBX.TG 

SHA256: 9f1f11a708d393e0a4109ae189bc64f1f3e312653dcf317a2bd406f18ffcc507 
MD5: 2915b3f8b703eb744fc54c81f4a9c67f  
Talos Rep: https://talosintelligence.com/talos_file_reputation?s=9f1f11a708d393e0a4109ae189bc64f1f3e312653dcf317a2bd406f18ffcc507  
Example Filename: e74d9994a37b2b4c693a76a580c3e8fe_1_Exe.exe  
Detection Name: Win.Worm.Coinminer::1201 

SHA256: 41f14d86bcaf8e949160ee2731802523e0c76fea87adf00ee7fe9567c3cec610 
MD5: 85bbddc502f7b10871621fd460243fbc  
Talos Rep: https://talosintelligence.com/talos_file_reputation?s=41f14d86bcaf8e949160ee2731802523e0c76fea87adf00ee7fe9567c3cec610  
Example Filename: 85bbddc502f7b10871621fd460243fbc.exe 
Detection Name: W32.41F14D86BC-100.SBX.TG 

SHA256: 96fa6a7714670823c83099ea01d24d6d3ae8fef027f01a4ddac14f123b1c9974  
MD5: aac3165ece2959f39ff98334618d10d9  
Talos Rep: https://talosintelligence.com/talos_file_reputation?s=96fa6a7714670823c83099ea01d24d6d3ae8fef027f01a4ddac14f123b1c9974 
Example Filename: 96fa6a7714670823c83099ea01d24d6d3ae8fef027f01a4ddac14f123b1c9974.exe  
Detection Name: W32.Injector:Gen.21ie.1201 

SHA256: 629ff05b7396cd0278ac345008b1e9246a6511da973e3e7eb630c5890758c15a 
MD5: 6692f8df8616472715273203b42e754d  
Talos Rep: https://talosintelligence.com/talos_file_reputation?s=629ff05b7396cd0278ac345008b1e9246a6511da973e3e7eb630c5890758c15a  
Example Filename: EasyAsVPNgo.exe 
Detection Name: W32.Proxy.27e2.1201

Cisco Talos Blog – ​Read More

Threat actors predominately exploited public-facing applications for initial access this quarter, with this tactic appearing in over 60 percent of Cisco Talos Incident Response (Talos IR) engagements – a notable increase from less than 10 percent last quarter. This spike is largely attributable to a wave of engagements involving ToolShell, an attack chain that targets on-premises Microsoft SharePoint servers through exploitation of vulnerabilities that were publicly disclosed in July. We also saw an increase in post-exploitation phishing campaigns launched from compromised valid accounts this quarter, a trend we noted last quarter, with threat actors using this technique to expand their attack both within the compromised organizations as well as to external partner entities.  

Ransomware incidents made up only approximately 20 percent of engagements this quarter, a decrease from 50 percent last quarter, despite ransomware remaining one of the most persistent threats to organizations. Talos IR responded to Warlock, Babuk, and Kraken ransomware variants for the first time, while also responding to previously seen families Qilin and LockBit. We observed an attack we attributed with moderate confidence to the threat actor that Microsoft tracks as China-based group Storm-2603 based on overlapping tactics, techniques, and procedures (TTPs). As part of their attack chain, the actors leveraged open-source digital forensics and incident response (DFIR) platform Velociraptor for persistence, a tool that has not been previously seen in ransomware attacks or associated with Storm-2603. We also responded to more Qilin ransomware engagements than last quarter, supporting our assessment from last quarter that the threat group is likely accelerating the cadence of their attacks.

ToolShell attacks underscore importance of robust segmentation and rapid patching 

As mentioned above, threat actors exploited public-facing applications for initial access in over 60 percent of engagements this quarter. Almost 40 percent of all engagements involved ToolShell activity, majorly contributing to this tactic’s rise in popularity.  

Starting in mid-July 2025, threat actors began actively exploiting two path traversal vulnerabilities affecting on-premises SharePoint servers: CVE-2025-53770 and CVE-2025-53771. These two vulnerabilities are related to CVE-2025-49704 and CVE-2025-49706, which had been previously featured in Microsoft Patch Tuesday updates in early July. One of the key features of the older vulnerabilities was that the adversary needed to be authenticated to obtain a valid signature by extracting the ValidationKey from memory or configuration. With the discovery of the newer vulnerabilities, attackers managed to eliminate the need to be authenticated to obtain a valid signature, resulting in unauthenticated remote code execution. 

This quarter’s ToolShell activity highlights the importance of network segmentation, as attackers demonstrated how they can exploit poorly segmented environments once a single server is compromised to move laterally within a targeted network. For example, in one engagement, the victim organization was impacted by ToolShell exploitation against a SharePoint server, then experienced a ransomware attack a few weeks later. In the latter attack, Talos IR analysis indicated the actors transferred credential stealing malware from the affected public-facing SharePoint server to a SharePoint database server on the victim’s internal network, demonstrating how they leveraged the trusted relationship between the two servers to expand their foothold. 

The wave of ToolShell attacks also shows how quickly threat actors mobilize when significant zero-day vulnerabilities are disclosed and/or proof-of-concepts appear. Active exploitation of the ToolShell vulnerabilities was first observed in the wild on July 18, a day before Microsoft issued its emergency advisory. Almost all Talos IR engagements responding to ToolShell activity kicked off within the following 10 days. Automated scanning enables attackers to rapidly discover and exploit vulnerable hosts while defenders race to test and deploy patches across diverse environments. Patching as soon as possible is key in narrowing that window of exposure, in addition to building safeguards through robust segmentation as mentioned above.

Post-exploitation phishing attacks from compromised accounts persist 

Consistent with findings from last quarter, threat actors continued to launch phishing campaigns after their initial compromise by leveraging compromised internal email accounts to expand their attack both within the compromised organization as well as externally to partner entities. This tactic appeared in a third of all engagements this quarter, an increase from last quarter’s 25 percent. Last quarter, we predominately saw this tactic used when phishing was also used for initial access. This quarter, however, we also saw it appear in engagements where other methods, such as valid accounts, were used for initial access. 

The follow-on phishing campaigns were primarily oriented towards credential harvesting. For example, in one engagement, the adversary used a compromised Microsoft Office 365 account to send almost 3,000 emails to internal and external partners. To evade detection, the adversary modified the email management rules to hide the sent phishing emails and any replies. Almost 30 employees of the targeted organization received the adversary’s phishing email and at least three clicked on the malicious credential harvesting link that was included; it is unknown how many users at external organizations were impacted. In another engagement, the adversary used a compromised email account to send internal phishing emails containing a link that directed to a credential harvesting page. The malicious site mimicked an Office 365 login page that was configured to redirect to the targeted organization’s legitimate login page upon the user entering their credentials, enhancing the attack’s legitimacy.   

Looking forward, as defenses against phishing attacks improve, adversaries are seeking ways to enhance these emails’ legitimacy, likely leading to the increased use of compromised accounts post-exploitation that we have observed recently. Defenders should seek to improve identification and protection capabilities against internal phishing campaigns, with actions such as providing stronger authentication methods for users’ email accounts, enhancing analysis of users’ email patterns and notifying on anomalies, and improving user awareness training.

Ransomware trends 

Ransomware incidents made up approximately 20 percent of engagements this quarter, a decrease from 50 percent last quarter, though we assess this dip is likely not indicative of any larger downward trend in the ransomware threat environment. Talos IR responded to Warlock, Babuk and Kraken ransomware variants for the first time, while also responding to previously seen families Qilin and LockBit.

Open-source DFIR tool Velociraptor adopted into ransomware toolkit  

We responded to a ransomware engagement this quarter that we assessed with moderate confidence was attributable to the Storm-2603 threat group based on overlapping TTPs, such as the deployment of both LockBit and Warlock ransomware. Storm-2603 is a suspected China-based threat actor that was first seen in July 2025 when they engaged in ToolShell activity. While LockBit is widely deployed by various ransomware actors, Warlock was first advertised in June 2025 and has since been heavily used by Storm-2603. Notably, we also observed evidence of Babuk ransomware files on the customer’s network in this engagement, which has not been previously deployed by Storm-2603 according to public reporting, though it failed to encrypt and only renamed files. The incident severely impacted the customer’s IT environment, including connected Operational Support Systems (OSS), a critical component of telecommunication infrastructure that allows for remote management and monitoring of day-to-day operations. 

We discovered the actors installed an older version of open-source DFIR platform Velociraptor on five servers to maintain persistence and launched the tool several times even after the host was isolated. Velociraptor is a legitimate tool that we have not previously observed being abused in ransomware attacks. It is a free product designed to help with investigations, data collection, and remediation during and after security incidents and it provides real-time or near real-time visibility into the activities occurring on monitored endpoints. The version of Velociraptor observed in this incident was outdated (version 0.73.4.0) and exposed to a privilege escalation vulnerability (CVE-2025-6264), which may have been leveraged for persistence as this vulnerability can lead to arbitrary command execution and endpoint takeover. The addition of this tool in the ransomware playbook is in line with findings from Talos’ 2024 Year in Review, which highlights the increasing variety of commercial and open-source products leveraged by threat actors.

Qilin ransomware operators likely accelerate pace of attacks    

We saw an increased number of Qilin ransomware engagements kick off this quarter compared to last quarter, when we encountered it for the first time. We predicted last quarter the group was accelerating their operational tempo, based on an increase in disclosures on their data leak site since February 2025. We observed Qilin operators use TTPs consistent with last quarter, including valid compromised credentials for initial access, a binary encryptor customized to the victim, and file transfer tool CyberDuck for exfiltration. In one Qilin engagement this quarter we were able to determine the adversaries’ dwell time as well, finding that the ransomware was executed two days after the attack first began. The steady increase in Qilin activity indicates it will very likely continue to be a top ransomware threat through at least the remainder of 2025, pending any disruption or intervention.

Targeting 

For the first time since we began documenting analysis of Talos IR engagements in 2021, public administration was the most targeted industry vertical this quarter. Though it hasn’t been the top targeted vertical before, it is often amongst the most seen, making this observation not entirely unexpected. Organizations within the public administration sector are attractive targets as they are often underfunded and still using legacy equipment. Additionally, the organizations targeted this quarter were largely local governments, which also typically oversee and support public school districts and county-run hospitals or clinics. As such, these entities often have access to sensitive data as well as a low downtime tolerance, making them attractive to financially motivated and espionage-focused threat groups, both of which we observed during these engagements.

Initial access 

As mentioned, the most observed means of gaining initial access this quarter was exploitation of public-facing applications, largely due to ToolShell activity. Other observed means of achieving initial access included phishing, valid accounts, and drive-by compromise.

Recommendations for addressing top security weaknesses

Implement detections to identify MFA abuse and strong MFA policies for impossible travel scenarios 

Almost a third of engagements this quarter involved multifactor authentication (MFA) abuse, including MFA bombing and MFA bypass — a slight decrease from approximately 40 percent last quarter. MFA bombing, also known as an MFA fatigue attack, involves an attacker repeatedly sending MFA requests to a user’s device, aiming to overwhelm them into inadvertently approving an unauthorized login attempt. MFA bypass encompasses a range of techniques leveraged by attackers to circumvent or disable MFA mechanisms and gain unauthorized access. Talos IR recommends defenders implement detections to identify when MFA has been bypassed, such as deploying products that use behavior analytics to identify new device logins and policies to generate alerts when detected.  

Talos IR also encountered numerous engagements this quarter that involved impossible travel scenarios, and recommended organizations implement strong MFA policies when these are detected. An example of an impossible travel scenario would be if a user logs into their account from New York, then the adversary logs into the same account three minutes later from Tokyo.   

Configure centralized logging capabilities across the environment  

Insufficient logging hindered investigation and response in approximately a third of engagements, a slight increase from 25 percent last quarter, due to issues such as log retention limitations, logs that were encrypted or deleted during attacks, and lack of logs due to disablement by the adversary. Understanding the full context and chain of events performed by an adversary on a targeted host is vital not only for remediation but also for enhancing defenses and addressing any system vulnerabilities for the future. Talos IR recommends that organizations implement a Security Information and Event Management (SIEM) solution for centralized logging. In the event an adversary deletes or modifies logs on the host, the SIEM will contain the original logs to support forensic investigation. 

Conduct robust patch management  

Finally, vulnerable/unpatched infrastructure was exploited in approximately 15 percent of engagements this quarter. Targeted infrastructure included unpatched development servers and unpatched SharePoint servers that remained vulnerable weeks after the ToolShell patches were released — we did not include SharePoint servers that were vulnerable before the release of the patches in this category. Exploitation of vulnerable infrastructure enabled adversaries’ lateral movement, emphasizing the importance of patch management.

Top-observed MITRE ATT&CK techniques

The table below represents the MITRE ATT&CK techniques observed in this quarter’s Talos IR engagements. Given that some techniques can fall under multiple tactics, we grouped them under the most relevant tactic in which they were leveraged. Please note this is not an exhaustive list.  

Key findings from the MITRE ATT&CK framework include:  

• Related to the internal phishing campaigns observed this quarter, we saw adversaries leveraging email hiding rules in numerous engagements, hiding certain inbound and outbound emails in the compromised user’s mailbox to evade detection. We also saw user execution of malicious links that directed to credential harvesting pages in these campaigns.
• We observed web shells deployed for persistence in the ToolShell engagements this quarter. The most observed web shell, “spinstall0.aspx”, was used to extract sensitive cryptographic keys from compromised servers.

Cisco Talos Blog – ​Read More

We’ve previously written about why neural networks are not the best choice for private conversations. Popular chatbots like ChatGPT, DeepSeek, and Gemini collect user data for training by default, so developers can see all our secrets: every chat you have with the chatbot is stored on company servers. This is precisely why it’s essential to understand what data each neural network collects, and how to set them up for maximum privacy.

In our previous post, we covered configuring ChatGPT’s privacy and security in abundant detail. Today, we examine the privacy settings in China’s answer to ChatGPT — DeepSeek. Curiously, unlike in ChatGPT, there aren’t that many at all.

What data DeepSeek collects

• All data from your interactions with the chatbot, images and videos included
• Details you provide in your account
• IP address and approximate location
• Information about your device: type, model, and operating system
• The browser you’re using
• Information about errors

What’s troubling is that the company doesn’t specify how long it keeps personal data, operating instead on the principle of “retain it as long as needed”. The privacy policy states that the data retention period varies depending on why the data is collected, yet no time limit is mentioned. Is this not another reason to avoid sharing sensitive information with these neural networks? After all, dataset leaks containing users’ personal data have become an everyday occurrence in the world of AI.

If you want to keep your IP address private while you work with DeepSeek, use a Kaspersky Security Cloud. Be wary of free VPN apps: threat actors frequently use them to create botnets (networks of compromised devices). Your smartphone or computer, and by extension, you yourself, could thus become unwitting accomplices in actual crimes.

Who gets your data

DeepSeek is a company under Chinese jurisdiction, so not only the developers but also Chinese law enforcement — as required by local laws — may have access to your chats. Researchers have also discovered that some of the data ends up on the servers of China Mobile — the country’s largest mobile carrier.

However, DeepSeek is hardly an outlier here: ChatGPT, Gemini, and other popular chatbots just as easily and casually share user data upon a request from law enforcement.

Disabling DeepSeek’s training on your data

The first thing to do — a now-standard step when setting up any chatbots — is to disable training on your data. Why could this pose a threat to your privacy? Sometimes, large language models (LLMs) can accidentally disclose real data from the training set to other users. This happens because neural networks don’t distinguish between confidential and non-confidential information. Whether it’s a name, an address, a password, a piece of code, or a photo of kittens — it makes little difference to the AI. Although DeepSeek’s developers claim to have taught the chatbot not to disclose personal data to other users, there’s no guarantee this will never happen. Furthermore, the risk of dataset leaks is always there.

The web-based version and the mobile app for DeepSeek have different settings, and the available options vary slightly. First of all, note that the web version only offers three interface languages: English, Chinese, and System. The System option is supposed to use the language set as the default in your browser or operating system. Unfortunately, this doesn’t always work reliably with all languages. Therefore, if you need the ability to switch DeepSeek’s interface to a different language, we recommend using the mobile app, which has no issues displaying the selected user interface language. It’s important to note that your choice of UI language doesn’t affect the language you use to communicate with DeepSeek. You can chat with the bot in any language it supports. The chatbot itself proudly claims to support more than 100 languages — from common to rare.

DeepSeek web version settings

To access the data management settings, open the left sidebar, click the three dots next to your name at the bottom, select Settings, and then navigate to the Data tab in the window that appears. We suggest you disable the option labeled Improve the model for everyone to reduce the likelihood that your chats with DeepSeek will end up in its training datasets. If you want the model to stop learning from the data you shared with it before turning off this option, you’ll need to email privacy@deepseek.com, and specify the exact data or chats.

Disabling DeepSeek training on your data in the web-based version

DeepSeek mobile app settings

In the DeepSeek mobile app, you also open the left sidebar, click the three dots next to your name at the bottom, and reveal the Settings menu. In the menu, open the Data controls section and turn off Improve the model for everyone.

Disabling DeepSeek training on your data in the app

Managing DeepSeek chats

All your chats with DeepSeek — both in the web version and in the mobile app — are collected in the left sidebar. You can rename any chat by giving it a descriptive title, share it with anyone by creating a public link, or delete a specific chat entirely.

Sharing DeepSeek chats

The ability to share a chat might seem extremely convenient, but remember that it poses risks to your privacy. Let’s say you used DeepSeek to plan a perfect vacation, and now you want to share the itinerary with your travel companions. You could certainly create a public link in DeepSeek and send it to your friends. However, anyone who gets hold of that link can read your plan and learn, among other things, that you’ll be away from home on specific dates. Are you sure this is what you want?

If you’re using the chatbot for confidential projects (which is not advisable in the first place, as it’s better to use a locally running version of DeepSeek for this kind of data, but more on this later), sharing the chat, even with a colleague, is definitely not a good idea. In the case of ChatGPT, similar shared chats were at one point indexed by search engines — allowing anyone to find and read them.

If you absolutely must send the content of a chat to someone else, it’s easier to copy it by clicking the designated button below the message in the chat window, and then to use a conventional method like email or a messaging app to send it, rather than share it with the entire world.

If, despite our warnings, you still wish to share your conversation via a public link, this is currently only possible in the web version of DeepSeek. To create a link to a chat, click the three dots next to the chat name in the left sidebar, select Share, and then, on the main chat board, check the boxes next to the messages you want to share, or check the Select all box at the bottom. After this, click Create public link.

Sharing DeepSeek chats in the web version

You can view all the chats you have shared and, if necessary, delete their public links in the web version, by going to Settings → Data → Shared links → Manage.

Managing shared DeepSeek chats in the web version

Deleting old DeepSeek chats

Why should you delete old DeepSeek chats? The fewer chats you have saved, the lower the risk that your confidential data could become accessible to unauthorized parties if your account is compromised, or if there’s a bug in the LLM itself. Unlike ChatGPT, DeepSeek doesn’t remember or use data from your past chats in new ones, so deleting them won’t impact your future use of the neural network.

However, you can resume a specific chat with DeepSeek at any time by selecting it in the sidebar. Therefore, before deleting a chat, consider whether you might need it again later.

To delete a specific chat: in the web version, click the three dots next to the chat in the left sidebar; in the mobile app, press and hold the chat name. In the window that appears, select Delete.

To delete your entire conversation history: in the web version, go to Settings → Data → Delete all chats → Delete all; in the application, go to Settings → Data controls → Delete all chats. Bear in mind that this only removes the chats from your account without deleting your data from DeepSeek’s servers.

If you want to save the results of your chats with DeepSeek, in the web version, first go to Settings → Data → Export data → Export. Wait for the archive to be prepared, and then download it. All data is exported in the JSON format. This feature is not available in the mobile app.

Managing your DeepSeek account

When you first access DeepSeek, you have two options: either sign up with your email and create a password, or log in with a Google account. From a security and privacy standpoint, the first option is better — especially if you create a strong, unique password for your account: you can use a tool like Kaspersky Password Manager to generate and safely store one.

You can subsequently log in with the same account in other browsers and on different devices. Your chat history will be accessible from any device linked to your account. So, if someone learns or steals your DeepSeek credentials, they’ll be able to review all your chats. Sadly, DeepSeek doesn’t yet support two-factor authentication or passkeys.

If you’ve even the slightest suspicion that your DeepSeek account credentials have been compromised, we recommend taking the following steps. Start by logging out of your account on all devices. In the web version, navigate to Settings → Profile → Log out of all devices → Log out. In the app, the path is Settings → Data controls → Log out of all devices. Next, you need to change your password, but DeepSeek doesn’t offer a direct path to do so once you’re logged in. To reset your password, go to the DeepSeek web version or mobile app, select the password login option, and click Forgot password?. DeepSeek will request your email address, send a verification code to that email, and allow you to reset the old password and create a new one.

Deploying DeepSeek locally

Privacy settings for the DeepSeek web version and mobile app are extremely limited and leave much to be desired. Fortunately, DeepSeek is an open-source language model. This means anyone can deploy the neural network locally on their computer. In this scenario, the AI won’t train on your data, and your information won’t end up on the company’s servers or with third parties. However, there’s a significant downside: when running the AI locally, you’ll be limited to the pre-trained model, and won’t be able to ask the chatbot to find up-to-date information online.

The simplest way to deploy DeepSeek locally is by using the LM Studio application. It allows you to work with models offline, and doesn’t collect any information from your chats with the AI. Download the application, click the search icon, and look for the model you need. The application will likely offer many different versions of the same model.

Searching LM Studio for DeepSeek models

These versions differ in the number of parameters, denoted by the letter B. The more parameters a model has, the more mathematical computations it can perform, and the better it performs; consequently, the more resources it requires to run smoothly. For comparison, a modern laptop with 16–32GB of RAM is sufficient for lighter models (7B–13B), but for the largest version, with 70 billion parameters, you’d need to own an entire data center.

LM Studio will alert you if the model is too heavy for your device.

LM Studio warning you that the model may be too large for your device

It’s important to understand that local AI use is not a panacea in terms of privacy and security. It doesn’t hurt to periodically check that LM Studio (or another similar application) is not connecting to external servers. For example, you can use the netstat command for that. If you’re not familiar with netstat, simply ask the chatbot to tell you about today’s news. If the chatbot is running locally as designed, the response definitely won’t include any current events.

Furthermore, you mustn’t forget about protecting the devices themselves: malware on your computer can intercept your data. Use Kaspersky Premium: it allows you to examine and block hidden connections, and will alert you to the presence of malicious software.

More on secure AI use:

• Privacy settings in ChatGPT
• How phishers and scammers use AI
• The pros and cons of AI-powered browsers
• Should you disable Microsoft Recall in 2025?
• Trojans masquerading as DeepSeek and Grok clients
• How AI can leak your private data

Kaspersky official blog – ​Read More