DeepSeek for Business
DeepSeek's models are strong and open, but its hosted app raises real data protection and security questions. What regulators and testers found, and who should use it.

DeepSeek is a capable and genuinely open family of AI models, and its hosted app is the part a European business should be careful with. Those are two different products, and most of the argument about whether DeepSeek is safe comes from treating them as one.
This review separates the two. It covers what DeepSeek is, where your data goes when you use the app, what regulators and independent testers have found, what running the models yourself changes, and who should and should not use it. It is written as a data governance review. It judges the service by where it stores data, what its own policy says and what has been measured, and nothing else.
Key takeaways
- DeepSeek publishes its models under the MIT licence, so you can download and run them yourself. That is a real strength.
- The hosted app and API are a separate matter. DeepSeek’s privacy policy says data is stored on servers in China, and European regulators have questioned whether that meets GDPR rules on transfers.
- In January 2025 Wiz Research found a DeepSeek database exposed without authentication, holding over a million lines of logs including chat history. It was secured promptly.
- A US government-backed evaluation found earlier DeepSeek models far easier to hijack as agents. That test pre-dates the V4 series.
- Running the open weights on your own systems removes the data transfer, but not the model’s behaviour.
- A safe default is to keep customer data, contracts and confidential material out of the hosted app, and to write down what staff may use it for.
What is DeepSeek?
DeepSeek is a Chinese AI company that builds large language models and offers them through a chat app, a paid programming interface and downloadable model files. It became widely known in early 2025 when its reasoning models performed close to leading Western models at a much lower reported training cost.
The current generation is the V4 series. The model card for DeepSeek-V4-Pro describes a preview version with 1.6 trillion parameters in total, of which 49 billion are active for each token, a context window of one million tokens, and a licence stated as MIT for both the repository and the weights. The technical report was published in April 2026.
Two things follow. First, this is a serious model, not a toy, and you should not dismiss it on origin. Second, “open” here means the weights are available to download. It does not describe the app on the company’s own servers, which is a different system with its own terms. The rest of this review keeps them apart.
Is DeepSeek Safe to use at work?
For some uses, yes. For others, not without a decision. The honest answer depends on three questions you can answer yourself.
What are you putting into it? A request to rewrite a paragraph of your own marketing copy carries little risk. A pasted customer list, a draft contract or an employee record carries a lot. The sensible line is the one in what is safe to paste, and it applies to every hosted AI tool, not only this one.
Where does the processing happen? If it happens on DeepSeek’s servers, the next section matters. If it happens on your own infrastructure or a European host you have vetted, it does not.
Who is accountable? If nobody owns the decision, staff will make it for you. A tool that nobody approved but everyone uses is shadow AI, and it is how most data incidents with AI start.
None of these is about who wrote the model. They are about data, jurisdiction and process, and they would apply to a tool from any country.
Where does your Data go when you use DeepSeek?
To servers that DeepSeek’s privacy policy says are located in the People’s Republic of China. That is the central fact, and press coverage of the policy, including from Tom’s Hardware, reports it consistently. The same coverage says the data collected goes beyond what you type and includes technical details about your device.
That matters for a European business because of data residency and the GDPR. Personal data can leave the EU only on a lawful basis, such as an adequacy decision or approved safeguards. If a member of staff pastes a customer’s personal details into the app, your business may be the one that has made an unlawful transfer, whether or not anyone intended it.
Be precise about what is and is not established. The policy’s wording on storage is clear. How the company handles the data after that, how long it keeps it and who can access it, are matters on which the policy is broad and regulators have asked for more. A broad policy is a reason for caution, not proof of misuse. The sensible response to uncertainty is to keep sensitive data out.
What did regulators do about DeepSeek?
Several European authorities acted in 2025, and the pattern is consistent.
In early 2025, Italy’s data protection authority, the Garante, ordered limits on the processing of Italian users’ data with immediate effect, after judging DeepSeek’s answers to its questions insufficient. Ireland’s Data Protection Commission wrote to the company asking for details of how it processes Irish users’ data, which is a request for information, not a ban.
In June 2025 the Berlin data protection commissioner reported the app to Apple and Google as illegal content. Her stated reason was that DeepSeek had not shown that German users’ data would be protected in China to a level equivalent to the EU’s, and that its transfers breached Article 46 of the GDPR. She had first asked DeepSeek to meet the transfer requirements or withdraw the app voluntarily, and it did not do so. Other authorities, including in Belgium, France and the Netherlands, opened their own enquiries.
I could not confirm from the sources available whether the app stores removed the app in Germany, or whether any authority has since issued a final decision. Treat the 2025 actions as a record of what regulators were concerned about, and check each authority’s current position before you decide. Regulatory positions move, and the one that applies to you is the one in your own country.
Has DeepSeek had a Security incident?
Yes, one that is well documented. On 29 January 2025 Wiz Research published that it had found a publicly accessible database belonging to DeepSeek. It was a ClickHouse database, reachable through two open ports on DeepSeek subdomains, and it required no authentication at all. Wiz described over a million lines of log streams containing chat history, secret keys and backend details, with full control over database operations available to anyone who found it.
Two points keep this in proportion. Wiz said it did not run intrusive queries beyond enumeration, and DeepSeek secured the exposure promptly once told. And an exposed database is a configuration mistake of a kind that has hit companies of every size and country.
It still matters for what it showed. At a point when millions of people were trying the app, chat histories were sitting in plain text on an open endpoint. That is a lesson about how much any provider should be trusted with sensitive text, and a reason to prefer a service whose controls you can assess. It is not evidence of what happens to data today.
How does DeepSeek hold up against attacks?
The best independent evidence is a 2025 evaluation by the US National Institute of Standards and Technology, through its AI standards and evaluation unit, known as CAISI. It tested the R1, R1-0528 and V3.1 models against several US models, and published its findings on 30 September 2025.
The headline results were these. Agents built on R1-0528, which NIST called DeepSeek’s most secure model in the test, were on average 12 times more likely than the evaluated US frontier models to follow malicious instructions, including in simulated cases that led to phishing emails, malware being run and login credentials being leaked. When a common jailbreaking technique was used, R1-0528 responded to 94% of overtly malicious requests, compared with 8% for the US reference models. The report also said the best US model outperformed the best DeepSeek model in the test across almost every benchmark.
Read this with care. It is a government agency of one country, evaluating a competitor’s models, so you should weigh the source. It tested older models, not the V4 series. And jailbreak resistance matters most for a public chatbot, while agent hijacking matters most if you let the model act on your behalf.
That second point is the one for a business to take away. If you connect any model to email, files or tools, it can be steered by instructions hidden in the content it reads, which is what prompt injection is. A model that is weaker against it needs more guardrails, and a person kept in the loop for anything it does.
The same evaluation reported that DeepSeek models echoed four times as many inaccurate and misleading Chinese Communist Party narratives as the US reference models. For drafting and summarising this will rarely be noticeable. For research on politics, news or geopolitics, check answers independently.
What does DeepSeek do well?
Quite a lot, which is why it deserves a fair review.
It is open. The MIT licence on the V4 weights is permissive. You can download the model, run it, adapt it and build on it, with fewer restrictions than many commercial licences. That is rare at this quality.
It reasons well for its class. The reasoning models were what first made the company’s name, and the NIST report itself acknowledges that the models are competitive, even while saying that the best US model was ahead.
It is built for efficiency. The mixture-of-experts design means only a fraction of the parameters is active for each token, which keeps running costs down for people who host it themselves.
It gives you an alternative. For a business that wants a model it can run under its own control, a strong open model is valuable. That is the same attraction as Mistral Large 4, which comes from a European company with a different set of trade-offs.
None of this is in dispute. The question is whether the strengths are worth the care that the hosted app requires, and whether you can capture the strengths without the exposure.
Does running DeepSeek yourself solve the problem?
It solves one problem and leaves the others.
What it solves: the data flow. If you run the weights on your own servers or on a European host you trust, your prompts never go to DeepSeek’s infrastructure. The transfer question, the storage question and the question of who else can read your chats largely fall away. That is the strongest argument for the open weights, and it is the same logic behind comparing cloud AI and on-premise AI generally.
What it does not solve:
- Model behaviour. The weaknesses NIST measured are properties of the model, not of where it runs. An agent built on it is still more exposed to hijacking than the alternatives tested.
- Effort and cost. A model this size needs serious hardware, or a hosting partner, plus someone to maintain it. For most small businesses that is not realistic, and the trade-offs are set out in local models compared with Claude.
- Supply chain. Download weights only from the official repository, and run them in an environment without unnecessary outbound access, as you would for any software from outside your organisation.
- Accuracy. Like every large language model, it can hallucinate, so the checking routine is the same.
For most small businesses, self-hosting is more than they need, and the practical question is simply whether to allow the app.
How does DeepSeek compare with other options?
The comparison depends on what matters to you, so here it is by question, without a score.
If your priority is data control, a model you run yourself, or a vendor with a European data processing agreement and clear storage terms, is a safer fit than the DeepSeek app. A European option such as Mistral Large 4 is one place to look, and the general choice is covered in ChatGPT and Claude compared.
If your priority is capability per unit of cost, DeepSeek’s open models are competitive, and self-hosting can be economical at scale, but the NIST evaluation found the best US model ahead on almost every benchmark it ran.
If your priority is safety when the model acts for you, the 2025 evidence favours the US models tested. Do not give any model more access than the task needs.
If your priority is compliance, the hosted app is the hardest to defend. A regulator who asks why you sent customer data to a service that European authorities had questioned will want to see a documented decision, and “a member of staff tried it” is not one. The wider picture is in the EU AI Act and the GDPR.
Who should use DeepSeek, and who should not?
Reasonable to use: a technically capable team that wants an open model and can host it under its own control; individuals trying it on public, non-sensitive material; developers experimenting with the weights in an isolated environment.
Use with a written rule: staff using the app for low-risk drafting, with the rule stating what may never go in. Put it in your AI use policy so it is a decision, not an accident.
Avoid: anyone handling customer data, health, financial, legal or employee records through the hosted app; any business that has to demonstrate GDPR compliance on a transfer and cannot; any set-up where an agent built on the model has access to email, files or payments without a person approving each action.
If staff are already using it, banning AI tools outright rarely works, because people move to personal accounts. A short, clear rule works better. If you are not sure what is already in use, an AI audit finds out.
Verdict
DeepSeek is a strong open model attached to a hosted service that a European business should treat with caution. Judge the two separately.
The weights are a legitimate option for a team that can run them itself. The hosted app is a poor fit for anything confidential, because of where it stores data, the regulatory questions that remain open and the security record so far. The 2025 evidence on model security is real but dated, and the V4 series has not had an equivalent public evaluation that I could find.
My advice is simple. Do not ban it by reflex and do not adopt it by default. Decide what may go into it, write that down, and keep personal and confidential data out of the app. If you want to set those rules with someone who has done it before, a scoping conversation is the place to start.
Sources
- NIST, “CAISI Evaluation of DeepSeek AI Models Finds Shortcomings and Risks” (30 September 2025): the US standards body’s evaluation of the R1, R1-0528 and V3.1 models, reporting that agents based on R1-0528 were on average 12 times more likely to be hijacked than the US models tested, a 94% response rate to malicious requests under a common jailbreak against 8% for US models, and four times as many echoed Chinese Communist Party narratives.
- Wiz Research, “Wiz Research Uncovers Exposed DeepSeek Database Leaking Sensitive Information, Including Chat History” (29 January 2025): a security firm’s account of an unauthenticated ClickHouse database holding over a million lines of logs, including chat history and secret keys, which DeepSeek secured promptly after disclosure.
- DeepSeek-V4-Pro model card, Hugging Face: the developer’s own description of a preview model with 1.6 trillion parameters, 49 billion active, a one-million-token context and an MIT licence on the repository and weights.
- Tom’s Hardware, “Chinese-made DeepSeek AI model records extensive online user data, stores it in China-based servers”: press coverage of DeepSeek’s privacy policy, which states that collected information is stored on servers in the People’s Republic of China.
- TechRadar, “Deepseek faces ban in Germany as privacy watchdog reports the app to Google and Apple as illegal content”: reporting on the Berlin data protection commissioner’s June 2025 action over transfers of German users’ data to China under Article 46 of the GDPR.
- Tom’s Guide, “DeepSeek AI banned in Italy as data privacy concerns pile up”: reporting on the Italian data protection authority’s order limiting the processing of Italian users’ data in early 2025.
FAQ
Questions we get asked
Is DeepSeek Safe to use for Business?
It depends on how you use it. The hosted app and API send what you type to servers that DeepSeek's privacy policy says are in China, and European regulators have questioned whether that meets GDPR rules on transfers. Running the open weights on your own infrastructure removes that data flow. Either way, keep customer data, contracts and anything confidential out of it until you have made a documented decision.
Where does DeepSeek store your Data?
DeepSeek's privacy policy says information it collects is stored on servers in the People's Republic of China. That covers what you type into the app and, according to press coverage of the policy, technical details about your device. If you run the open weights yourself, your prompts stay on your own systems, because there is no DeepSeek server in the loop.
Is DeepSeek banned in Europe?
Not across the whole of Europe. Italy's data protection authority ordered limits on processing Italian users' data in early 2025, and Berlin's data protection commissioner reported the app to Apple and Google as illegal content in June 2025, saying DeepSeek had not shown that German users' data would be protected to an EU-equivalent level. Other authorities opened enquiries. Check each regulator's current position, because it can change.
Is DeepSeek open Source?
Its models are open weights under the MIT licence, which means you can download and run them. The DeepSeek-V4-Pro model card says the repository and the model weights are under the MIT licence. Open weights are not the same as a hosted service. The downloadable model and the DeepSeek app are different things with different data risks.
Can DeepSeek leak your Data?
Any hosted service can, and DeepSeek has already had one exposure. In January 2025 Wiz Research found a DeepSeek database reachable without authentication, holding over a million lines of logs including chat history and secret keys. DeepSeek secured it promptly after being told. That is a past incident, not proof of a current fault, but it is relevant to how much you trust the service with sensitive material.
Is the DeepSeek Model less Secure than ChatGPT or Claude?
In one government-backed test, yes. The US National Institute of Standards and Technology evaluated earlier DeepSeek models in 2025 and found that agents built on R1-0528 were on average 12 times more likely than the US models tested to follow malicious instructions. That test pre-dates the V4 series, so it is evidence about earlier models, not a verdict on the latest.
Does DeepSeek censor its Answers?
The same NIST evaluation reported that DeepSeek models echoed four times as many inaccurate and misleading Chinese Communist Party narratives as the US reference models. For most business tasks, such as drafting or summarising, that will rarely matter. For research, news analysis or anything touching politics, treat answers as needing independent checking.
Should staff be allowed to use the DeepSeek app?
Only for non-sensitive tasks, and only if your AI use policy says so. The bigger risk is staff using it on personal accounts without telling anyone, which is shadow AI. Decide, write it down and tell people. A clear rule that allows a safe use is followed more often than a blanket ban that people quietly ignore.
Ready to Put This to Work?
Tell us where your team is with AI and we will tell you honestly what would make the biggest difference.

