DeepSeekA consultant tries DeepSeek because a colleague told her it was as good as the expensive tools, and free. She pastes in a client's situation (names, the contract value, a bank account, the dispute) and asks for a summary of possible strategies. The answer is genuinely good.
What she does not see is where her prompt has just gone. According to DeepSeek's own privacy policy, the text she typed, the file she uploaded and the chat history are collected, processed and stored on servers in the People's Republic of China: a jurisdiction with no EU adequacy decision, whose laws can compel a company to hand data to the state.
The model performed beautifully. That was never the question. The question is where her client's data now lives, and who can ask to see it.
Is DeepSeek safe for business data?
For the hosted app and API, no, and the reason has little to do with how capable the model is. DeepSeek's privacy policy states plainly that it stores personal data in China. For an EU organisation, that is an international transfer to a country without an adequacy decision, and the policy cites no Standard Contractual Clauses to make it lawful. That is exactly why Italy's regulator blocked it within days and Germany later reported the app to Apple and Google.
But there is an important fork in the road. DeepSeek publishes its model weights under a permissive licence. Run those weights on your own infrastructure, or with an EU provider that hosts them, and no prompt ever travels to China. The privacy problem is not the model but the hosted service. The decision that matters is where DeepSeek runs.
How DeepSeek handles your data
| Hosted app / API | Self-hosted open weights | |
|---|---|---|
| Where prompts go | Servers in China | Your own or EU infrastructure |
| EU adequacy decision | None | Not applicable: data stays put |
| Data processing agreement | Not offered for consumers | You control it |
| Used to improve the model | Yes (default) | No, unless you choose to |
| Suitable for regulated data | No | Yes, with proper controls |
The hosted DeepSeek app (mobile and web) and the cloud API send everything to China: what you type, uploaded files, photos and chat history, plus device and network data. The policy also allows data to be shared with authorities "to comply with applicable law, legal process or government requests". Deleting a chat from your account does not guarantee it is deleted from the servers.
The open weights are a different product entirely. Licensed for commercial use and self-hosting, they let an organisation run DeepSeek's model where the data already lives. This route turns DeepSeek from a compliance problem into an ordinary choice of supplier. The largest models do need serious hardware, though, so most teams would use a managed EU or US host for the weights.
Five risks, all about where the data ends up
1. Prompts stored in a country without an adequacy decision
This is the core issue. Everything typed into the hosted app lands in China, outside the GDPR's framework for transfers. For client, patient or employee data, that is a risk no setting in the app can remove.
2. The risk of state access
Chinese law can compel companies to assist the state's intelligence services. DeepSeek's policy already provides for sharing data in response to "government requests". For sensitive commercial or personal data, that is a structural risk, not a hypothetical one.
3. Onward transfer without consent
South Korea's data protection authority found in April 2025 that DeepSeek had transferred users' data, including prompt content, to Volcano Engine, a ByteDance affiliate, without consent or disclosure. It ordered the prompt data to be deleted.
4. The open-database exposure
In January 2025 the security firm Wiz found a DeepSeek database that was publicly accessible without authentication and exposed over a million log lines, including chat history in plain text and API keys. DeepSeek secured it quickly once notified, but it showed how immature the operations behind a fast-growing service can be.
5. The "it's just a chatbot" framing
Because DeepSeek looks and behaves like any other assistant, employees treat it like one and paste the same sensitive details they might (still unwisely) paste into a managed tool. The interface hides the jurisdiction.
Why this is more than a privacy question
Under the GDPR, sending personal data of someone in the EU to China without a valid transfer mechanism is in itself a breach of the rules on international transfers (Articles 44-49), even before anything is misused. The organisation whose employee pastes the data is the controller making that transfer, so the liability lies with the employer as well as with DeepSeek.
Fines can reach €20 million or 4% of global annual turnover, whichever is higher. And regulators have shown they act fast: Italy's ban took effect within two days of the Garante raising its concerns. Beyond the regulator, telling a client that their confidential details were stored on servers in China is a conversation few businesses want to have.
What regulators found at DeepSeek
January 2025: Wiz exposes an open DeepSeek database
Wiz Research found a ClickHouse database belonging to DeepSeek that was open to the internet without authentication and leaked over a million log lines, including chat history and secret keys. DeepSeek secured it within about 30 minutes of being notified. Source: Wiz Research, January 2025.
January 2025: Italy's Garante imposes an immediate ban
Italy's data protection authority ordered an immediate and definitive restriction on the processing of Italian users' data, after finding that, under DeepSeek's policy, data was stored in China in breach of the GDPR's security obligations. DeepSeek was removed from Italian app stores. Source: Garante / Euronews, January 2025.
April 2025: South Korea finds unauthorised transfers
South Korea's Personal Information Protection Commission (PIPC) found that DeepSeek had transferred Korean users' data abroad without consent, including prompt content sent to a ByteDance affiliate, and ordered it to put this right. Source: PIPC / Korea Times, April 2025.
June 2025: Germany refers DeepSeek to Apple and Google
Berlin's data protection authority formally reported the DeepSeek apps to Apple and Google as illegal content under the Digital Services Act, citing unlawful data transfers to China and breaches of the GDPR's transfer safeguards. Source: Berlin data protection authority / TechCrunch, June 2025.
The pattern is not a single flaw. It is a service whose default is to move your data somewhere EU law cannot follow it.
Where DeepSeek lands before anyone reads a privacy policy
None of this requires an employee to research Chinese data law or wait for a regulator's ruling. In BeeSensible's catalogue of over 1,000 AI tools, DeepSeek has a Critical risk score, the band reserved for exactly this kind of unresolved jurisdiction and retention risk. Nothing in the catalogue is quietly approved along the way: every tool starts with the status "No decision yet" and keeps it until someone in the organisation sets it to Allowed, Limited use or Not allowed. A Critical score is not background noise in a report. It is the first thing the admin sees when DeepSeek turns up to be assessed.
