Skip to content
Aventail™ investigation into UK government AI dependence featured in Computer Weekly.Read the story
Aventail™
All guides

Is DeepSeek safe for work?

15 min read

Is DeepSeek safe for work? A dated, sourced verdict for the DeepSeek app, the API and the open weights, with the privacy terms, the bans and the 2025 leak.

The DeepSeek app, website and API are not safe for company data. All three are run from China by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., which stores what you send on servers in China and publishes no business plan, admin controls or audit log. The open-weight DeepSeek models are a separate case: on hardware you control they send nothing to DeepSeek. The short answer, as of 15 November 2026:

What you are using

Safe for company data?

Why

DeepSeek app and website (chat.deepseek.com, iOS, Android)

No

Prompts, files and chat history are stored in China. Content can be used to improve the models unless each person opts out. No admin, no SSO, no record of what was pasted. Several governments ban it on official devices.

DeepSeek API (api.deepseek.com)

No

Same company, same account system, same Chinese jurisdiction. The API terms state no retention period, no zero-retention option and no training commitment.

Open weights on your own hardware

Yes, after testing

The weights are MIT-licensed files. Run offline, they have no route to DeepSeek. What remains is the model's behaviour: weaker resistance to malicious instructions and politically shaped output in tests of earlier versions.

Open weights through another host

Depends on the host

The data terms are the host's. Amazon Bedrock, for example, states that model providers cannot see customer prompts. A host you have not vetted is a new third party.

Everything below comes from DeepSeek's own policies, regulator statements and named security research, read on 15 November 2026 and listed at the end. Check the review date before you quote the tables in a policy.

What DeepSeek does with your data, by product

App and website

API

Open weights, self-hosted

Who runs it

Hangzhou DeepSeek Artificial Intelligence Co., Ltd., registered in China

The same company; the Open Platform and the chat service share one account

You, or the host you choose

Where data is processed and stored

"We directly collect, process and store your Personal Data in People's Republic of China"

The privacy policy applies to the developer's account. The API terms name no other region

Wherever the machine is

Used to improve or train models

Yes, by default. The terms allow use of inputs and outputs to "improve the Services or the underlying technologies"

Not stated in the Open Platform terms. There is no clause that excludes it

No. Nothing is sent

How to opt out

Turn off "Improve the model for everyone" in settings, or email privacy@deepseek.com

No setting or clause published

Not needed

How long data is kept

"For as long as you have an account". No fixed period for chats

Not published

Your own retention

Admin controls, SSO, audit log

None published. Personal accounts only

API keys and balance only. No SSO, SCIM or audit log documented

Whatever your own stack provides

Certifications

None named in the policies, site or API docs we read

None named

Not applicable

Governing law

People's Republic of China. Disputes go to the court where DeepSeek is registered

People's Republic of China

Your own contracts

Price

Free

deepseek-flash: $0.15 to $0.30 per million input tokens, $0.60 to $1.20 per million output tokens. deepseek-v4-pro: $0.66 to $1.32 in, $1.98 to $3.96 out. Off-peak is half the peak rate

MIT licence, no fee. You pay for hardware

The global English privacy policy was last updated on 10 February 2026, the terms of use on 27 March 2026 and the Open Platform terms took effect on 29 April 2026. DeepSeek also has a desktop agent, DeepSeek Harness, in public preview. We did not review its data flows for this page.

What the privacy policy and terms mean for your staff

Everything typed is collected. The policy lists "text input, voice input, prompt, uploaded files, photos, feedback, chat history" as User Input. It also collects device model, operating system, IP address, device identifiers and approximate location.

It stays in China and can be shared inside the group. DeepSeek says entities in its corporate group process user data for "storage, content delivery, security, research and development, foundation model training and optimization". It may share data with "law enforcement agencies, public authorities" to "comply with applicable law, legal process or government requests". The applicable law is Chinese law.

Training is on until each person turns it off. The opt-out is a personal setting. Nobody at your company can set it for staff or check that it was set. South Korea's regulator found in April 2025 that DeepSeek had launched without any opt-out and added one on 17 March 2025 during the investigation.

DeepSeek itself says not to send sensitive data. The policy states the services "are not designed or intended to process sensitive Personal Data" and that "you should not provide sensitive Personal Data". The EU has no adequacy decision for China, which is the point the Berlin regulator acted on.

The API is the same company with thinner terms. The Open Platform terms give you ownership of outputs and make you the controller for your end users' data. They say nothing about how long API inputs are kept or whether they are used for training, and they disclaim that the service will be "uninterrupted, timely, secure or free from error". There is no data processing agreement on the pages we read.

Regulator actions and government bans

Date

Who

What happened

30 January 2025

Italy, data protection authority (Garante)

Ordered an urgent limitation on processing Italian users' data by Hangzhou DeepSeek and Beijing DeepSeek and opened an investigation. The companies had said European law did not apply to them.

January to early February 2025

United States

The Navy told personnel not to use DeepSeek "in any capacity". NASA blocked it on agency devices and networks. The Pentagon's IT agency blocked the website. Texas banned it on state devices on 31 January.

By 3 February 2025

Taiwan

The Ministry of Digital Affairs banned government agencies, public schools and state-owned enterprises from using it.

Early February 2025

Australia

PSPF Direction 001-2025 found that DeepSeek products pose an unacceptable security risk to the Commonwealth Government, and they were banned from government devices. Victoria's matching guideline tells public bodies to prevent installation and remove existing instances. It excludes open-source models deployed locally on a government system.

February to April 2025

South Korea, PIPC

New app downloads were suspended in mid-February. The 24 April findings: personal data sent to China and the US without consent, and user input sent to Beijing Volcano Engine Technology, which DeepSeek stopped on 10 April. The app returned at the end of April.

27 June 2025

Germany, Berlin data protection commissioner

Reported the DeepSeek apps to Apple and Google as illegal content under the Digital Services Act, after DeepSeek did not comply with a 6 May demand. The finding: transfers to China breach Article 46(1) GDPR. Her 2025 annual report says both stores refused to block the app and that she will file a complaint.

9 and 10 July 2025

Czech Republic

The government banned DeepSeek products, websites and the API in public administration. The cyber agency NÚKIB rated the threat "High". Both texts exempt DeepSeek's open-source models "deployed locally" with no ability to reach DeepSeek's servers.

Late July 2025

Netherlands

The government confirmed civil servants may not use the app. The privacy regulator had opened an investigation in January.

9 September 2026

South Korea, PIPC

A further corrective recommendation after a follow-up inspection, as recorded by Digital Policy Alert.

10 November 2026

United Kingdom, DWP

The Department for Work and Pensions' AI security policy, updated on this date, states: "Users must not attempt to access DeepSeek AI on DWP devices." We found no UK-wide government ban.

Two things stand out. The actions are about the hosted service and where its data goes. And two of the governments that banned it, the Czech Republic and Victoria, wrote an exception for locally deployed open-source models into the text.

The January 2025 exposed database

On 29 January 2025 Wiz Research reported a ClickHouse database belonging to DeepSeek that was open to the internet with no authentication, at oauth2callback.deepseek.com:9000 and dev.deepseek.com:9000. It held more than a million log lines, dating from 6 January 2025, with chat history in plain text, API secrets and backend details. Wiz said the exposure "allowed for full database control". DeepSeek closed it promptly after Wiz reported it.

It was fixed quickly. It still shows the state of the hosted service in January 2025: user chats sat in a log store that anyone could query.

Why the open weights are a different risk

DeepSeek publishes its model weights under the MIT licence on Hugging Face. The current releases are DeepSeek-V4-Pro and DeepSeek-V4.1-Flash, and the V4.1 Flash model card gives vLLM and SGLang commands for serving it. A model file is not a service. When you run it on a machine with no route to DeepSeek, the privacy policy, the Chinese servers and the bans above do not apply, because no data goes anywhere.

Three risks remain.

The model's behaviour. NIST's CAISI tested DeepSeek R1, R1-0528 and V3.1 in 2025. Agents built on the most secure of them were on average 12 times more likely than US frontier models to follow malicious instructions that hijacked the task. With a common jailbreak the model answered 94% of overtly malicious requests, against 8% for the US reference models. It repeated four times as many inaccurate Chinese Communist Party narratives. CrowdStrike tested the raw R1 weights in November 2025 and found that politically sensitive words in the prompt raised the rate of code with severe vulnerabilities by up to 50%. Neither study covers the V4 models, so run your own tests on the version you deploy.

Hardware. DeepSeek describes V4.1 Flash as a 552B-parameter model. That is server hardware. The models that fit a laptop or workstation are the older distilled R1 sizes, from 1.5B to 70B on Ollama.

The name on the label. On Ollama's library on 15 November 2026, deepseek-v4.1-flash and deepseek-v4-pro are listed as cloud models. A cloud model runs on Ollama's servers. If your reason for choosing open weights is that data must stay on your machines, check that the model is a local download before anyone uses it.

Through another host, read that host's terms. AWS states that in Amazon Bedrock model providers "don't have access to Amazon Bedrock logs or to customer prompts and completions". That removes DeepSeek from the data path and puts AWS in it, under your AWS contract.

The red list: what never goes into the DeepSeek app or API

For most tools in this series the red list is a short set of exceptions. For DeepSeek's hosted service it is everything that is not already public: personal data, source code, credentials, contracts, financial results, privileged or regulated material, and internal documents of any kind.

If staff want DeepSeek for its quality or price, give them the open weights on approved hardware.

Settings to change today

If staff have used the app or website for work:

  1. Turn off "Improve the model for everyone" in settings. It covers new use only.

  2. Delete the chat history, then the account. DeepSeek says a deleted account cannot be recovered.

  3. Rotate any password, key or token that was pasted into a chat.

  4. Record what was shared, for your data protection lead to assess.

If developers have used the API:

  1. Revoke the keys in the DeepSeek platform and remove them from repositories and CI.

  2. Check coding tools for a DeepSeek base URL. Our guides to Claude Code with DeepSeek and OpenCode with DeepSeek show where that setting lives.

On the network and devices:

  1. Block chat.deepseek.com, platform.deepseek.com and api.deepseek.com at the proxy or DNS filter.

  2. Block the mobile app in MDM.

If you self-host the weights:

  1. Download from the official deepseek-ai organisation on Hugging Face and record the version.

  2. Serve the model on a machine or network segment with no outbound route to deepseek.com.

  3. Test tool use and jailbreak resistance with your own prompts before a wider rollout.

What to put in the acceptable use policy

Three lines cover DeepSeek. The DeepSeek app, website and API are not approved for any company data. DeepSeek open-weight models are approved only on named, company-controlled infrastructure. Anyone who has already pasted company data into the app reports it without blame. Our guide to shadow AI and acceptable use policies has the full outline.

Run the open weights on your own machines

The safe way to use a DeepSeek model at work is to run the weights yourself and make sure requests cannot leave. Doing that for one developer is a download and a command. Doing it for a team needs a rule that each machine follows and a record that shows it did.

Aventail™ Link is an inference engine and gateway that runs on each machine. It serves GGUF language models, and Aventail™ Control accepts GGUF uploads, so an open-weight model your organisation has approved can be distributed to devices from one place. Aventail™ Control holds the routing rules and Link evaluates them on the device. A rule can permit only On device. If the local model cannot serve the request, the request fails with an error. It is not sent to a disallowed destination. Prompts and responses do not pass through the hosted Control service; it receives management, device and routing events.

Aventail™ Connect then points supported apps (Claude Desktop, Claude Code, OpenCode and Codex) at Link, with no endpoint or key to enter. Routing policies are part of Control for Enterprise. That gives you two lanes:

  • The DeepSeek app, website and API: blocked.

  • An approved open-weight model on your own machines: routed by Aventail™, with a Requests view in Control that shows where each request ran.

The free plan covers three connected devices, including business use. Check a model's size against the device's memory before you choose it, and test it on one device first, as Aventail™'s deployment guide advises. For the wider picture, see self-hosted AI.

Frequently asked questions

Does DeepSeek send your data to China?
The app, website and API do. DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China. A DeepSeek model that you run on your own machine sends nothing.

Is it safe to run DeepSeek locally?
For data, yes: a local model has no connection to DeepSeek, and the Czech and Victorian government bans both exempt locally deployed open-source models. For output, treat it like any open model. Tests of earlier versions found it easier to jailbreak and hijack than US models, so review what it writes and limit what an agent built on it can do.

Is the DeepSeek API safer than the app?
No. It is the same company and the same jurisdiction. The API terms publish no retention period, no zero-retention option and no commitment about training.

Does DeepSeek train on my chats?
By default, yes. The terms allow DeepSeek to use inputs and outputs to improve its services and technology. Each user can opt out by turning off "Improve the model for everyone".

Is DeepSeek banned in the UK?
We found no UK-wide ban. The Department for Work and Pensions bans it on its own devices. Italy, Australia, Taiwan, the Czech Republic, the Netherlands, South Korea and several US agencies and states have restricted the hosted service since January 2025.

Is DeepSeek safer through AWS or another cloud?
The risk changes owner. On Amazon Bedrock, AWS says the model provider cannot see prompts or completions, so the data terms are your AWS terms. Check each host's own statement.

Has DeepSeek had a data breach?
In January 2025 Wiz Research found a DeepSeek database open to the internet with more than a million log lines, including chat history and API secrets. DeepSeek secured it after the report.

Sources

All read on 15 November 2026. DeepSeek serves its privacy policy and terms of use by region, and our connection was redirected to the Korean edition (updated 6 May 2026). The global English editions quoted here were read from Internet Archive captures dated 12 November 2026.

  1. DeepSeek, Privacy Policy (last update 10 February 2026): https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html

  2. DeepSeek, Terms of Use (last update 27 March 2026): https://cdn.deepseek.com/policies/en-US/deepseek-terms-of-use.html

  3. DeepSeek, Open Platform Terms of Service (effective 29 April 2026): https://cdn.deepseek.com/policies/en-US/deepseek-open-platform-terms-of-service.html

  4. DeepSeek, Model Mechanism and Training Methods: https://cdn.deepseek.com/policies/en-US/model-algorithm-disclosure.html

  5. DeepSeek API docs, Models and pricing: https://api-docs.deepseek.com/quick_start/pricing

  6. DeepSeek API docs, Change log: https://api-docs.deepseek.com/updates

  7. DeepSeek, home page and Transparency Center: https://www.deepseek.com/en and https://www.deepseek.com/en/transparency/

  8. DeepSeek, V4.1 Flash announcement: https://www.deepseek.com/en/news/deepseek-v4-1-flash/

  9. Hugging Face, deepseek-ai organisation and DeepSeek-V4.1-Flash-1017 model card: https://huggingface.co/deepseek-ai

  10. Ollama library, DeepSeek models: https://ollama.com/search?q=deepseek

  11. Garante per la protezione dei dati personali, press release (30 January 2025): https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/10097450

  12. PIPC (South Korea), Status examination results of DeepSeek service (24 April 2025): https://pipc.go.kr/eng/user/ltn/new/noticeDetail.do?bbsId=BBSMSTR_000000000001&nttId=2819

  13. Digital Policy Alert, PIPC investigation into DeepSeek (entry of 9 September 2026): https://digitalpolicyalert.org/change/21475-personal-information-protection-commissions-investigation-into-deepseek-over-alleged-provision-of-rights-exercise-guidance-exclusively-in-chinese

  14. Berlin Commissioner for Data Protection, press release (27 June 2025): https://www.datenschutz-berlin.de/pressemitteilung/berliner-datenschutzbeauftragte-meldet-ki-app-deepseek-in-deutschland-bei-apple-und-google-als-rechtswidrigen-inhalt/

  15. Berlin Commissioner for Data Protection, Annual report 2025: https://www.datenschutz-berlin.de/jahresbericht-2025/

  16. NÚKIB (Czech Republic), Warning regarding certain products of DeepSeek (10 July 2025): https://nukib.gov.cz/en/infoservis-en/news/2280-nukib-issues-a-warning-regarding-certain-products-of-the-company-deepseek

  17. Victorian Government, Direction on the use of DeepSeek products (updated 14 February 2025): https://www.vic.gov.au/direction-use-deepseek-products-applications-and-web-services

  18. Office of the Texas Governor, ban on Chinese AI and social media apps (31 January 2025): https://gov.texas.gov/news/post/governor-abbott-announces-ban-on-chinese-ai-social-media-apps

  19. TechCrunch, The countries and agencies that have banned DeepSeek (3 February 2025): https://techcrunch.com/2025/02/03/deepseek-the-countries-and-agencies-that-have-banned-the-ai-companys-tech

  20. Reuters via Insurance Journal, Governments, regulators increase scrutiny of DeepSeek (7 January 2026): https://amp.insurancejournal.com/news/international/2026/01/07/853376.htm

  21. UK Department for Work and Pensions, Artificial Intelligence Security Policy (updated 10 November 2026): https://www.gov.uk/government/publications/dwp-procurement-security-policies-and-standards/artificial-intelligence-security-policy

  22. NIST, CAISI evaluation of DeepSeek AI models (30 September 2025): https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks

  23. CrowdStrike, Hidden vulnerabilities in AI-coded software (20 November 2025): https://www.crowdstrike.com/en-us/blog/crowdstrike-researchers-identify-hidden-vulnerabilities-ai-coded-software/

  24. Wiz Research, Exposed DeepSeek database (29 January 2025): https://www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak

  25. AWS, Data protection in Amazon Bedrock: https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html

  26. AWS, DeepSeek in Amazon Bedrock: https://aws.amazon.com/bedrock/deepseek/

  27. Aventail™, Connect overview: https://docs.aventail.co.uk/docs/connect/overview

  28. Aventail™, Routing: https://docs.aventail.co.uk/docs/fleet/routing

  29. Aventail™, Manage models and Compatibility and versions: https://docs.aventail.co.uk/docs/guides/models and https://docs.aventail.co.uk/docs/introduction/versions-and-supported-models

  30. Aventail™, Plan a deployment: https://docs.aventail.co.uk/docs/guides/deployment

  31. Aventail™, Pricing: https://aventail.co.uk/#pricing

We use optional cookies and tracking to identify visiting companies, load our contact form, measure Google Ads download clicks and understand page use with PostHog on our homepage, Connect and How it works pages, Ollama and LM Studio comparison guides and campaign pages. Cookie policy