The sudden rise of DeepSeek has raised questions of data origin, data destination, and the security of the new AI model.

The sudden rise of DeepSeek has raised concerns and questions, especially about the origin and destination of the training data, as well as the security of the data.

For those returning from a short holiday away from the news, DeepSeek is a new player on the Artificial Intelligence (AI) field. The Chinese startup has certainly taken the app stores by storm: In just a week after the launch it topped the charts as the most downloaded free app in the US. This caused an upset on the stock markets that cost nVidia and Oracle shareholders a lot of money.

DeepSeek has been called an open-source project, however this technically is not true because only the model’s outputs and certain aspects are publicly accessible. This makes it qualify as an open-weight model. Anyway, the important difference is that the underlying training data and code necessary for full reproduction of the models are not fully disclosed.

And it’s the data that pose a concern to many. OpenAI has accused DeepSeek of using its ChatGPT model to train DeepSeek’s AI chatbot, which triggered quite some memes. If only because OpenAI previously suffered accusations of using data that was not its own in order to train ChatGPT.

Authorities have started to ask questions as well. The Italian privacy regulator GPDP has asked DeepSeek to provide information about the data it processes in the chatbot, and its training data.  Because it sees a risk to the privacy of millions of Italian citizens, GDPD has demanded DeepSeek answers within 20 days questions about:

*   Which personal data is collected
*   The origin of the data
*   Purpose for the collection
*   Whether the data is stored on servers in China

According to the Italian press agency ANSA, DeepSeek disappeared on January 29, 2025 from Google and Apple’s app stores in Italy.

And if all that isn’t scary enough, researchers at Wiz have found a publicly accessible database belonging to DeepSeek.

> “This database contained a significant volume of chat history, backend data and sensitive information, including log streams, API Secrets, and operational details. “

The database was not just accessible and readable, it was also open to control and privilege escalation within the DeepSeek environment. No authentication was required, so anybody that stumbled over the database was able to run queries to retrieve sensitive logs and actual plaintext chat messages, and even to steal plaintext passwords and local files.

Needless to say, this oversight put DeepSeek and its users at risk.

We have said this before and we’ll probably have to repeat it numerous times, but the need for fast developments in this field is creating privacy risks that we have never seen before, simply because security is an afterthought for the developers. So, no matter which AI chatbot you prefer, always be mindful of the information you feed it: It may find its way to unexpected and undesirable places.

**We don’t just report on threats – we help safeguard your entire digital identity**

Cybersecurity risks should never spread beyond a headline. Protect your—and your family’s—personal information by using identity protection.

 The DeepSeek controversy: Authorities ask where does the data come from and how safe is it? 

Amateurish financial scams are common across Africa, and Namibia's influential former first lady, Monica Geingos, has emerged as a particularly effective host body for these messages.

Source: Sipa USA via Alamy Stock Photo

Scammers are using the likeness of Namibia's former first lady to trick people into falling for investment scams.

For some time now, Monica Geingos — first lady of Namibia from 2015 until her husband's passing early last year — has been warning Namibians about criminals using her likeness to perpetrate cheap financial scams. In a video posted online last week, Geingos reported an "uptick" in accounts impersonating her across social media platforms, with some using manipulated videos. Dark Reading has reached out to representatives of Mrs. Geingos for more information.

**Namibian Impersonation Scams**

Across Africa, as much as anywhere else on the globe, impersonation of political leaders, celebrities, and other generally influential figures feature prominently in investment scams.

Rachael Shitanda, vice chair of East Africa at the Africa Information & Communication Technologies Alliance (AfICTA), recalls a wealth of similar incidents affecting her home country of Kenya. "There were cases where fraudsters created fake social media accounts for Transport Cabinet Secretary Kipchumba Murkomen, promising job opportunities in exchange for 'processing fees.' These types of scams often target vulnerable job seekers. In another instance, fraudsters impersonated Deputy President Rigathi Gachagua in a fake investment scheme. They used doctored videos and photos to make the scam more convincing, leading several Kenyans to fall victim."

Geingos' combination of regional influence, having been voted one of Namibia's 12 most influential people in 2012 and one of Africa's 100 most influential women in 2020, and financial experience — she is a former employee of Namibia's Stock Exchange (NSX) and chairman of the board of eBank Namibia — make her a particularly effective vessel for these kinds of impersonation scams.

Source: Monica Geingos, The 3rd First Lady of Namibia on Facebook

Typically, scammers have impersonated Geingos on social media to promote fake foreign exchange (forex) trading platforms, or other forms of get-rich-quick schemes. Recently, Gengos said in her video, "Somebody received a video call, purportedly of me, convincing them to invest in some kind of weird forex \[foreign exchange\] schemes." These sorts of videos appear to be using real footage, dubbed over with fake audio intended to sound like a recording of her voice.

**Combating Online Scams in Africa**

These tricks seem to be working, as evidenced by the rising attacks against Geingos. "Something that has surprised \[my team\] of late is how ubiquitous it's become across social media platforms," she has said. And in one case, Geingos told the BBC, a woman who fell victim to one of these scams was swindled out of her retirement savings.

To Shitanda, there are a few main reasons why cyber scams might be particularly effective in Africa. First, law enforcement bodies have limited resources to fight rapidly evolving cyber threats. Plus, the criminals themselves operate across national borders, making them harder to track down and apprehend.

That said, Namibia's government has made some efforts to push back. Its state-owned Bank of Namibia founded a cybersecurity council aimed at combating online fraud in 2022, and a national cybercrime bill is currently undergoing a redrafting by the country's Ministry of Information and Communication Technology (MICT). Meanwhile, last summer, the International Criminal Police Organization (INTERPOL) concluded a major operation called First Light that, among targets in 60 other countries, busted up a scam ring in Namibia's capital city, Windhoek. Authorities seized hundreds of computers and mobile phones, and freed 88 children "who were forced into conducting scams."

Law enforcement aside, though, scammers are aided by more intractable core issues with how people engage with the Web. Digital literacy is lagging in some parts of the continent, Shitanda explains, and citizens of her country have a culture of trust in what they see: "People often believe messages from public figures without questioning their authenticity."

**About the Author**

Nate Nelson is a writer based in New York City. He formerly worked as a reporter at Threatpost, and wrote "Malicious Life," an award-winning Top 20 tech podcast on Apple and Spotify. Outside of Dark Reading, he also co-hosts "The Industrial Security Podcast."

Fake Videos of Former First Lady Scam Namibians

China-based DeepSeek has exploded in popularity, drawing greater scrutiny. Case in point: Security researchers found more than 1 million records, including user data and API keys, in an open database.

The Chinese generative artificial intelligence platform DeepSeek has had a meteoric rise this week, stoking rivalries and generating market pressure for United States–based AI companies, which in turn has invited scrutiny of the service. Amid the hype, researchers from the cloud security firm Wiz published findings on Wednesday that show that DeepSeek left one of its critical databases exposed on the internet, leaking system logs, user prompt submissions, and even users’ API authentication tokens—totaling more than 1 million records—to anyone who came across the database.

DeepSeek is a relatively new company and has been virtually unreachable to press and other organizations this week. In turn, the company did not immediately respond to WIRED’s request for comment about the exposure. The Wiz researchers say that they themselves were unsure about how to disclose their findings to the company and simply sent information about the discovery on Wednesday to every DeepSeek email address and LinkedIn profile they could find or guess. The researchers have yet to receive a reply, but within a half hour of their mass contact attempt, the database they found was locked down and became inaccessible to unauthorized users. It is unclear whether any malicious actors or authorized parties accessed or downloaded any of the data.

“The fact that mistakes happen is correct, but this is a dramatic mistake, because the effort level is very low and the access level that we got is very high,” Ami Luttwak, the CTO of Wiz tells WIRED. “I would say that it means that the service is not mature to be used with any sensitive data at all.”

Exposed databases that are accessible to anyone on the open internet are a long-standing problem that institutions and cloud providers have slowly worked to address. But the Wiz researchers note that the DeepSeek database they found was visible almost immediately with minimal scanning or probing.

“Usually when we find this kind of exposure, it’s in some neglected service that takes us hours to find—hours of scanning,” says Nir Ohfeld, the head of vulnerability research at Wiz. But this time, “here it was at the front door.” Ohfeld adds that the “technical difficulty of this vulnerability is the bare minimum.”

The researchers say that the trove they found appears to have been a type of open source database typically used for server analytics called a ClickHouse database. And the exposed information supported this, given that there were log files that contained the routes or paths users had taken through DeepSeek’s systems, the users’ prompts and other interactions with the service, and the API keys they had used to authenticate. The prompts the researchers saw were all in Chinese, but they note that it is possible the database also contained prompts in other languages. The researchers say they did the absolute minimum assessment needed to confirm their findings without unnecessarily compromising user privacy, but they speculate that it may even have been possible for a malicious actor to use such deep access to the database to move laterally into other DeepSeek systems and execute code in other parts of the company’s infrastructure.

“It's pretty shocking to build an AI model and leave the backdoor wide open from a security perspective,” says independent security researcher Jeremiah Fowler, who was not involved in the Wiz research but specializes in discovering exposed databases. “This type of operational data and the ability for anyone with an internet connection to access it and then manipulate it is a major risk to the organization and users.”

DeepSeek’s systems are seemingly designed to be very similar to OpenAI’s, the researchers told WIRED on Wednesday, perhaps to make it easier for new customers to transition to using DeepSeek without difficulty. The entire DeepSeek infrastructure appears to mimic OpenAI’s, they say, down to details like the format of the API keys.

The Wiz researchers say they don’t know if anyone else found the exposed database before they did, but it wouldn’t be surprising, given how simple it was to discover. Fowler, the independent researcher, also notes that the vulnerable database would have “definitely” been found quickly—if it wasn’t already—whether by other researchers or bad actors.

“I think this is a wake-up call for the wave of AI products and services we will see in the near future and how seriously they take cybersecurity,” he says.

DeepSeek has made a global impact over the past week, with millions of people flocking to the service and pushing it to the top of Apple’s and Google’s app stores. The resulting shock waves have wiped billions from the stock prices of US-based AI companies and spooked executives at firms across the country. On Wednesday, sources at OpenAI told the Financial Times that it was looking into DeepSeek’s alleged use of ChatGPT outputs to train its models.

At the same time, DeepSeek has increasingly drawn the attention of lawmakers and regulators around the world, who have started to ask questions about the company’s privacy policies, the impact of its censorship, and whether its Chinese ownership provides national security concerns.

Italy’s data protection regulator sent DeepSeek a series of questions asking about where it obtained its training data, if people’s personal information was included in this, and the firm’s legal grounding for using this information. As WIRED Italy reported, the DeepSeek app appeared to be unavailable to download within the country following the questions being sent.

DeepSeek’s Chinese connections also appear to be raising security concerns. At the end of last week, according to CNBC reporting, the US Navy issued an alert to its personnel warning them not to use DeepSeek’s services “in any capacity.” The email said Navy members of staff should not download, install, or use the model, and raised concerns of “potential security and ethical” issues.

However, despite the hype, the exposed data shows that almost all technologies relying on cloud-hosted databases can be vulnerable through simple security lapses. “AI is the new frontier in everything related to technology and cybersecurity,” Wiz’s Ohfeld says, “and still we see the same old vulnerabilities like databases left open on the internet.”

Exposed DeepSeek Database Revealed Chat Prompts and Internal Data

Yet another spinoff of the infamous DDoS botnet is exploiting a known vulnerability in active attacks, while its threat actors are promoting it on Telegram for other attackers to use as well, in a DDoS-as-a-service model.

Source: Kirill Ivanov via Alamy Stock Photo

Yet another Mirai botnet variant is making the rounds, this time offering distributed denial-of-service (DDoS) as-a-service by exploiting flaws in Mitel SIP phones. It also features a unique capability to communicate with attacker command-and-control (C2).

Researchers at the Akamai Security Intelligence and Response Team (SIRT) identified the variant of the infamous botnet, dubbed Aquabot, that actively exploits CVE-2024-41710, a command-injection vulnerability that affects various Mitel models that are used in corporate environments, they revealed in a blog post published Jan. 29. The vulnerability relies on an input sanitization flaw, and exploitation can lead to root access of the device, SIRT researchers Kyle Lefton and Larry Cashdollar wrote in the post.

The variant is the third version of Aquabot (Akamai calls it Aquabotv3) to appear on the scene; the first version was built off the Mirai framework with the ultimate goal of DDoS, discovered in November 2023, and it was first reported by Antiy Labs. The second version of the bot "tacked on concealment and persistence mechanisms, such as preventing device shutdown and restart" that remain present in v3, the researchers wrote.

The new variant is distinct from the previous versions for a couple of reasons, the researchers said. One is a unique feature appearing first in Aquabotv3: a function named "report\_kill" that reports back to the C2 when a kill signal is caught on the infected device. So far, however, researchers have not seen any response to the function from the attacker C2.

Related:Unpatched Zyxel CPE Zero-Day Pummeled by Cyberattackers

Another notable aspect of v3 of Aquabot is that the threat actors behind it have been advertising the botnet as DDoS as-a-service through platforms such as Telegram. The bot is advertised under several different names — including Cursinq Firewall, The Eye Services, and The Eye Botnet — offering Layer 4 and Layer 7 DDoS, the researchers noted.

**Active Exploitation of Mitel Phone Security Flaw**

Akamai SIRT detected exploit attempts targeting CVE-2024-41710 through its global network of honeypots in early January using a payload almost identical to a proof-of-concept (PoC) developed and released on GitHub in mid-August by Packetlabs' researcher Kyle Burns.

Burns discovered that the Mitel 6869i SIP phone, firmware version 6.3.0.1020, failed to sanitize user-supplied input properly, with multiple endpoints vulnerable to the flaw. His PoC demonstrated that an attacker could smuggle in entries otherwise blocked by the application's sanitization checks by sending a specially crafted HTTP POST request.

Related:Super Bowl LIX Could Be a Magnet for Cyberattacks

The exploitation activity that Akamai SIRT observed delivered a payload that attempts to fetch and execute a shell script called :bin.sh, which will in turn fetch and execute Mirai malware on the target system, the researchers wrote. The malware has support for a variety of different architectures, including x86 and ARM.

"Based on our analysis of the malware samples, we determined that this is a version of the Aquabot Mirai variant," specifically the latest evolution of the malware, Aquabotv3, the researchers wrote in the post.

In addition to being used in DDoS attacks, threat actors also are hawking Aquabot for DDoS-as-a-service, though they are trying to disguise the activity as "purely testing" for DDoS mitigation. However, the same domain featured in the ad promoting testing is actively spreading Mirai malware, the researchers noted.

"Threat actors will claim it's just a \[proof of concept\] or something educational, but a deeper analysis shows that they are in fact advertising DDoS as a service, or the owners are boasting about running their own botnet on Telegram," they wrote in the post.

**Mirai Botnet Remains Key Conduit for DDoS**

As the majority of botnets responsible for DDoS attacks are based on Mirai, "they predominantly target Internet of Things (IoT) devices, which makes spreading the malware relatively easy to do," the researchers noted in the post. Indeed, a recent wave of global DDoS attacks were attributed to Mirai botnet spinoffs, demonstrating that attackers aiming to leverage Mirai show no signs of slowing down.

Related:Apple Patches Actively Exploited Zero-Day Vulnerability

That's likely because "the \[return on investment\] of Mirai for an aspiring botnet author is high," because it's not only one of the most successful botnet families in the world, it's also one of the more simple ones to modify, the researchers noted.

Moreover, many IoT devices often lack proper security features, are at the end of service, or are left with default configurations and passwords either from neglect or lack of knowledge about the dangers, making them low-hanging fruit for Mirai and its variants, the researchers wrote.

No matter what an attacker's intentions are, the researchers recommended that organizations take action to secure IoT devices through discovery or changing default credentials to protect against DDoS threats.

"Many of these botnets rely on common password libraries for authentication," they wrote in the post. "Find out where your known IoT devices are, and check for rogue ones, too. Check the login credentials and change them if they are default or easy to guess."

Akamai SIRT also included a list of indicators of compromise (IoCs) as well as Snort and Yara rules in the post to aid defenders.

**About the Author**

Elizabeth Montalbano is a freelance writer, journalist, and therapeutic writing mentor with more than 25 years of professional experience. Her areas of expertise include technology, business, and culture. Elizabeth previously lived and worked as a full-time journalist in Phoenix, San Francisco, and New York City; she currently resides in a village on the southwest coast of Portugal. In her free time, she enjoys surfing, hiking with her dogs, traveling, playing music, yoga, and cooking.

Mirai Variant 'Aquabot' Exploits Mitel Device Flaws

A team of security researchers from Georgia Institute of Technology and Ruhr University Bochum has demonstrated two new side-channel attacks targeting Apple silicon that could be exploited to leak sensitive information from web browsers like Safari and Google Chrome.
The attacks have been codenamed Data Speculation Attacks via Load Address Prediction on Apple Silicon (SLAP) and Breaking the

New SLAP & FLOP Attacks Expose Apple M-Series Chips to Speculative Execution Exploits

Apple has released a host of security updates for iOS, iPadOS, Mac, Apple Watch, and Apple TV. Update as soon as you can.

Apple has released a host of security updates across many devices, including for a zero-day bug which is being actively exploited in iOS.

Apple said:

> “A malicious application may be able to elevate privileges. Apple is aware of a report that this issue may have been actively exploited against versions of iOS before iOS 17.2.”

Devices affected are those that run:

*   iPhone XS and later
*   iPad Pro 13-inch, iPad Pro 12.9-inch 3rd generation and later, iPad Pro 11-inch 1st generation and later, iPad Air 3rd generation and later, iPad 7th generation and later, and iPad mini 5th generation and later
*   macOS Sequoia
*   Apple Watch Series 6 and later
*   All models of Apple TV HD and Apple TV 4K

If you use any of these then you should install updates as soon as you can. To check if you’re using the latest software version, go to **Settings** (or **System Settings**) > **General** \> **Software Update**. It’s also worth turning on Automatic Updates if you haven’t already, which you can do on the same screen.

**Technical details about the zero-day**

The zero-day vulnerability patched in this update is tracked as CVE-2025-24085. It is described as a use after free (UAF) issue in Apple’s Core Media framework that would allow an attacker to elevate privileges.

The Core Media framework handles multimedia applications like photos, videos, and real-time communication applications. UAF is a type of vulnerability that is the result of the incorrect use of dynamic memory during a program’s operation. If, after freeing a memory location, a program does not clear the pointer to that memory, an attacker can use the error to manipulate the program. Referencing memory after it has been freed can cause a program to crash, use unexpected values, or execute code. In this case, successful exploitation could provide a malicious app with privileges on the affected device that it shouldn’t have.

**We don’t just report on phone security—we provide it**

Cybersecurity risks should never spread beyond a headline. Keep threats off your mobile devices by downloading Malwarebytes for iOS, and Malwarebytes for Android today.

 Apple users: Update your devices now to patch zero-day vulnerability 

Apple has released software updates to address several security flaws across its portfolio, including a zero-day vulnerability that it said has been exploited in the wild.
The vulnerability, tracked as CVE-2025-24085, has been described as a use-after-free bug in the Core Media component that could permit a malicious application already installed on a device to elevate privileges.
"Apple is

Apple Patches Actively Exploited Zero-Day Affecting iPhones, Macs, and More

The Apple iOS 18.3 update fixes 28 other vulnerabilities identified by the tech company, though there is little information on them.

Source: Shahid Jamil via Alamy Stock Photo

NEWS BRIEF

In its latest security update for users, Apple has released a patch for a zero-day vulnerability tracked as CVE-2025-24085 (no CVSS score assigned yet).

The vulnerability, not yet added to the National Vulnerability Database (NVD), can be found in iOS, iPadOS, macOS, tvOS, watchOS, and visionOS. As a privileged escalation security flaw, it is located in Apple's Core Media framework. The bug is being actively exploited in the wild.

The Core Media framework, according to Apple, is "the media pipeline used by AVFoundation and other high-level media frameworks found on Apple platforms." It allows users to process media samples as well as manage queues of media data.

This patch comes in the form of iOS 18.3, which fixes 28 other vulnerabilities as well. Apple has yet to divulge many details about any of the issues that have been patched by this update, likely to prevent attackers from exploiting them before users can apply the necessary fix.

Impacted devices from this bug include:

*   Apple Watch Series 6 and later
    

*   Apple TV HD and Apple TV 4K (all models)
    
*   iPad Pro 13-inch, iPad Pro 12.9-inch 3rd generation and later, iPad Pro 11-inch 1st generation and later, iPad Air 3rd generation and later, iPad 7th generation and later, and iPad mini 5th generation and later
    

Though the tech giant has disclosed that "Apple is aware of a report that this issue may have been actively exploited against versions of iOS before iOS 17.2," it has not published any details of the attacks nor attributed its discovery to a researcher.

**About the Author**

Skilled writer and editor covering cybersecurity for Dark Reading.

Apple Patches Actively Exploited Zero-Day Vulnerability

One of the largest data breaches in history was apparently twice as impactful as previously thought, with PII belonging to hundreds of millions of people sitting in the hands of cybercriminals.

Source: Pavel Kapish via Alamy Stock Photo

New evidence suggests that more than half of the US population was touched by the ransomware attack(s) against UnitedHealth subsidiary Change Healthcare.

One of the largest data breaches ever recorded struck Change Healthcare last year. Change's technology services reach hundreds of vendors and laboratories, thousands of hospitals, tens of thousands of pharmacies, and hundreds of thousands of physicians and dentists, including "nearly all government and commercial payers," according to company documentation. These services necessarily sweep up loads of patients' personally identifying information (PII), which ended up in the hands of multiple ransomware actors.

Later last year, it was reported that the incident affected around 100 million Americans. Now, UnitedHealth has updated that number to approximately 190 million. A company spokesperson confirmed this in an emailed statement to Dark Reading, adding, "the vast majority of those people have already been provided individual or substitute notice."

**Change's Changing Cyberattack Story**

Last February, in concert, pharmacies around the US experienced significant delays to prescription orders. Behind it all was Change Healthcare, which processes billions of transactions every year, representing trillions of dollars worth of medical claims.

Related:Actively Exploited Fortinet Zero-Day Gives Attackers Super-Admin Privileges

The company first stated that it had suffered a nation-state cyber intrusion. In fact, it was a regular old ransomware attack (for which it would later pay a whopping $22 million ransom). And this wasn't the only crucial detail it got wrong.

In June, Change Healthcare finally sent out notices of data compromise, revealing that affected customers totaled around 100 million. On Friday, however, UnitedHealth Group publicly adjusted that figure to include 90 million more.

In its updated online notice of data breach, the company admitted that hackers may have obtained a variety of personally identifying information (PII) about patients and guarantors, including their first and last names, dates of birth, phone numbers, home addresses, and email addresses. Social security numbers, it noted, were only lost in "rare instances," and in an email to Dark Reading, a spokesperson claimed that Change Healthcare "has not seen electronic medical record databases appear in the data during the analysis."

The spokesperson also emphasized that "Change Healthcare is not aware of any misuse of individuals' information as a result of this incident."

However, Paul Bischoff, consumer privacy advocate at Comparitech, warns that "every press release I ever see about a data release says 'there's no evidence that your information has been abused or misused in any way.' But, obviously, they're not really looking for that for those instances of abuse, and they would never know if it actually happened. And the people that it happens to can't attribute the identity theft that they're suffering back to the data breach that caused it."

Related:For $50, Cyberattackers Can Use GhostGPT to Write Malicious Code

**When Data Breach Disclosures Go Wrong**

The Securities and Exchange Commission (SEC) data breach disclosure rules require that publicly traded companies disclose "material" cybersecurity incidents within four days of becoming alerted to them. The same rule applies to material updates to breach disclosures, such as when an attack is found to have affected nearly twice as many victims as once thought.

Despite these rules, companies have managed to take extensive time in investigating and addressing critical aspects of their breaches. For instance, it took Change Healthcare four months to notify customers of its incident, nine months to admit that 100 million people were affected, and nearly a year to update that figure to 190 million.

Bischoff hesitates, though, before suggesting that what's needed is even stricter regulation. "It's a complicated subject, because it gets to a point where you put such a burden on companies. Companies are also victims in these situations, so I don't want to penalize them for reporting things incorrectly," he says.

Related:The Case for Proactive, Scalable Data Protection

At the same time, he adds, "What we see a lot is that these companies take way too long to finish their investigations and notify victims. Sometimes it's up to a year or more before we're notified that people's data is out there on the Dark Web, being used for who knows what. And that's after they're most likely to get hit with identity fraud, and other sorts of fraud, because cybercriminals want that information when it's as fresh as possible. That's when it's most valuable. So I think we do need more strict standards about the timeliness of these notifications."

**About the Author**

Nate Nelson is a writer based in New York City. He formerly worked as a reporter at Threatpost, and wrote "Malicious Life," an award-winning Top 20 tech podcast on Apple and Spotify. Outside of Dark Reading, he also co-hosts "The Industrial Security Podcast."

Change Healthcare Breach Impact Doubles to 190M People

Amid ongoing fears over TikTok, Chinese generative AI platform DeepSeek says it’s sending heaps of US user data straight to its home country, potentially setting the stage for greater scrutiny.

The United States’ recent regulatory action against the Chinese-owned social video platform TikTok prompted mass migration to another Chinese app, the social platform “Rednote.” Now, a generative artificial intelligence platform from the Chinese developer DeepSeek is exploding in popularity, posing a potential threat to US AI dominance and offering the latest evidence that moratoriums like the TikTok ban will not stop Americans from using Chinese-owned digital services.

DeepSeek, an AI research lab created by a prominent Chinese hedge fund, recently gained popularity after releasing its latest open source generative AI model that easily competes with top US platforms like those developed by OpenAI. However, to help avoid US sanctions on hardware and software, DeepSeek created some clever workarounds when building its models. On Monday, DeepSeek’s creators limited new sign-ups after claiming the app had been overrun with a “large-scale malicious attack.”

While DeepSeek has several AI models, some of which can be downloaded and run locally on your laptop, the majority of people will likely access the service through its iOS or Android apps or its web chat interface. Like with other generative AI models, you can ask it questions and get answers; it can search the web; or it can alternatively use a reasoning model to elaborate on answers.

DeepSeek, which does not appear to have established a communications department or press contact yet, did not return a request for comment from WIRED about its user data protections and the extent to which it prioritizes data privacy initiatives.

As people clamor to test out the AI platform, though, the demand brings into focus how the Chinese startup collects user data and sends it home. Users have already reported several examples of DeepSeek censoring content that is critical of China or its policies. The AI setup appears to collect a lot of information—including all your chat messages—and send it back to China. In many ways, it’s likely sending more data back to China than TikTok has in recent years, since the social media company moved to US cloud hosting to try to deflect US security concerns

“It shouldn’t take a panic over Chinese AI to remind people that most companies in the business set the terms for how they use your private data” says John Scott-Railton, a senior researcher at the University of Toronto’s Citizen Lab. “And that when you use their services, you’re doing work for them, not the other way around.”

**What DeepSeek Collects About You**

To be clear, DeepSeek is sending your data to China. The English-language DeepSeek privacy policy, which lays out how the company handles user data, is unequivocal: “We store the information we collect in secure servers located in the People's Republic of China.”

In other words, all the conversations and questions you send to DeepSeek, along with the answers that it generates, are being sent to China or can be. DeepSeek’s privacy policies also outline the information it collects about you, which falls into three sweeping categories: information that you share with DeepSeek, information that it automatically collects, and information that it can get from other sources.

The first of these areas includes “user input,” a broad category likely to cover your chats with DeepSeek via its app or website. “We may collect your text or audio input, prompt, uploaded files, feedback, chat history, or other content that you provide to our model and Services,” the privacy policy states. Within DeepSeek’s settings, it is possible to delete your chat history. On mobile, go to the left-hand navigation bar, tap your account name at the bottom of the menu to open settings, and then click “Delete all chats.”

This collection is similar to that of other generative AI platforms that take in user prompts to answer questions. OpenAI’s ChatGPT, for example, has been criticized for its data collection although the company has increased the ways data can be deleted over time. Regardless of these types of protections, privacy advocates emphasize that you should not disclose any sensitive or personal information to AI chat bots.

“I would not input personal or private data in any such an AI assistant,” says Lukasz Olejnik, independent researcher and consultant, affiliated with King's College London Institute for AI. Olejnik notes, though, that if you install models like DeepSeek’s locally and run them on your computer, you can interact with them privately without your data going to the company that made them. Additionally, AI search company Perplexity says it has added DeepSeek to its platforms but claims it is hosting the model in US and EU data centers.

Other personal information that goes to DeepSeek includes data that you use to set up your account, including your email address, phone number, date of birth, username, and more. Likewise, if you get in touch with the company, you’ll be sharing information with it.

Bart Willemsen, a VP analyst focusing on international privacy at Gartner, says that, generally, the construction and operations of generative AI models is not transparent to consumers and other groups. People don’t know exactly how they work or the exact data they have been built upon. For individuals, DeepSeek is largely free, although it has costs for developers using its APIs. “So what do we pay with? What do we usually pay with: data, knowledge, content, information,” Willemsen says.

As with all digital platforms—from websites to apps—there can also be a large amount of data that is collected automatically and silently when you use the services. DeepSeek says it will collect information about what device you are using, your operating system, IP address, and information such as crash reports. It can also record your “keystroke patterns or rhythms,” a type of data more widely collected in software built for character-based languages. Additionally, if you purchase DeepSeek’s premium services, the platform will collect that information. It also uses cookies and other tracking technology to “measure and analyze how you use our services.”

A WIRED review of the DeepSeek website's underlying activity shows the company also appears to send data to Baidu Tongji, Chinese tech giant Baidu's popular web analytics tool, as well as Volces, a Chinese cloud infrastructure firm. In a social media post, Sean O'Brien, founder of Yale Law School's Privacy Lab, said that DeepSeek is also sending “basic” network data and “device profile” to TikTok owner ByteDance “and its intermediaries.

The final category of information DeepSeek reserves the right to collect is data from other sources. If you create a DeepSeek account using Google or Apple sign-on, for instance, it will receive some information from those companies. Advertisers also share information with DeepSeek, its policies say, and this can include “mobile identifiers for advertising, hashed email addresses and phone numbers, and cookie identifiers, which we use to help match you and your actions outside of the service.”

**How DeepSeek Uses Information**

Huge volumes of data may flow to China from DeepSeek’s international user base, but the company still has power over how it uses the information. DeepSeek’s privacy policy says the company will use data in many typical ways, including keeping its service running, enforcing its terms and conditions, and making improvements.

Crucially, though, the company’s privacy policy suggests that it may harness user prompts in developing new models. The company will “review, improve, and develop the service, including by monitoring interactions and usage across your devices, analyzing how people are using it, and by training and improving our technology,” its policies say.

DeepSeek’s privacy policy also says the company will also use information to “comply with \[its\] legal obligations”—a blanket clause many companies include in their policies. DeepSeek’s privacy policy says data can be accessed by its “corporate group,” and it will share information with law enforcement agencies, public authorities, and more when it is required to do so.

While all companies have legal obligations, those based in China do have notable responsibilities. Over the past decade, Chinese officials have passed a series of cybersecurity and privacy laws meant to allow state officials to demand data from tech companies. One 2017 law, for instance, says that organizations and citizens should “cooperate with national intelligence efforts.”

These laws, alongside growing trade tensions between the US and China and other geopolitical factors, fueled security fears about TikTok. The app could harvest huge amounts of data and send it back to China, those in favor of the TikTok ban argued, and the app could also be used to push Chinese propaganda. (TikTok has denied sending US user data to China’s government.) Meanwhile, several DeepSeek users have already pointed out that the platform does not provide answers for questions about the 1989 Tiananmen Square massacre, and it answers some questions in ways that sound like propaganda.

Willemsen says that, compared to users on a social media platform like TikTok, people messaging with a generative AI system are more actively engaged and the content can feel more personal. In short, any influence could be larger. “Risks of subliminal content alteration, conversation direction steering, in active engagement ought by that logic to lead to more concern, not less,” he says, “especially given how the inner workings of the model are widely unknown, its thresholds, borders, controls, censorship rules, and intent/personae largely left unscrutinized, and it being already so popular in its infancy stage.”

Olejnik, of King's College London, says that while the TikTok ban was a specific situation, US law makers or those in other countries could act again on a similar premise. “We can’t rule out that 2025 will bring an expansion: direct action against AI firms,” Olejnik says. “Of course, data collection may again be named as the reason.”

_Updated 5:27 pm EST, January 27, 2025: Added additional details about the DeepSeek website's activity._

_Updated 10:05 am EST, January 29, 2025: Added additional details about DeepSeek's network activity._

Tag

#apple