Is DeepSeek Safe or Banned in India? A Review of the Bans, Data Risks and Security Tests
Short answer: DeepSeek is not banned in India for ordinary users. The app and website remain available, and the government has said privacy concerns can be handled by hosting DeepSeek's open models on Indian servers. The Finance Ministry did tell its own staff in early 2025 to avoid AI tools, including DeepSeek and ChatGPT, on official devices.
Whether it is "safe" depends on how you use it. The official app and API store data on servers in China under DeepSeek's privacy policy. Independent security tests found DeepSeek's R1 model far easier to jailbreak than leading US models, and a US government evaluation found it more vulnerable to agent hijacking. Running DeepSeek's open-weight models on your own hardware, or through a provider hosting them outside China, removes the data-location issue but not the weaker safety guardrails.
We reviewed 11 public sources, including Reuters reporting on government actions, DeepSeek's privacy policy as analysed by Tom's Hardware, security tests by Cisco, Wiz, Palo Alto Networks and Enkrypt AI, and the US NIST evaluation, to answer three questions: is DeepSeek banned in India, what are the real risks, and how can Indian users and developers use it more safely.
Key findings
- No general ban in India. India has restricted official use by Finance Ministry staff, according to Reuters, but has not blocked the app for the public. The IT Minister said in January 2025 that DeepSeek models would be hosted on Indian servers.
- At least nine countries acted against it. Reuters lists actions by Italy, Australia, South Korea, Taiwan, Germany, the Czech Republic, the Netherlands, France and the United States, mostly limited to government devices or app stores.
- Data goes to China when you use the official app. DeepSeek's privacy policy says it stores personal data on servers in China, and it collects device information, IP address, prompts and uploaded files, according to Tom's Hardware.
- Guardrails are weak. Cisco's researchers reported a 100% attack success rate against DeepSeek R1 on 50 harmful prompts, compared with 26% for OpenAI's o1-preview.
- Self-hosting changes the picture. Open-weight models run on your own servers do not send data to DeepSeek, but they keep the model's safety weaknesses.
How we reviewed the evidence
We separated three questions that are often mixed together: legal status (bans and restrictions), data privacy (where your information goes) and model safety (how the model behaves under attack). For each, we used primary documents or reporting from established outlets, and we note where security tests were run by companies with commercial interests in AI security.
Is DeepSeek banned in India?
No. Here is what the Indian government has actually done, based on the sources we reviewed:
- Finance Ministry advisory (dated 29 January 2025). Reuters reported that the ministry told employees to avoid AI tools and apps such as ChatGPT and DeepSeek on office computers and devices, citing risks to the confidentiality of government data and documents. This is an internal advisory, not a public ban, and it covers ChatGPT too.
- Hosting on Indian servers (30 January 2025). IT Minister Ashwini Vaishnaw said DeepSeek's models would be hosted on Indian servers under the IndiaAI Mission, according to TechCrunch, adding that "privacy issues can be addressed by hosting open source models on Indian servers".
The context matters. India has banned more than 300 China-linked apps since 2020, including TikTok, according to TechCrunch. DeepSeek has not been added to that list, and the government's chosen approach so far is to host the open models locally rather than block them.
Which countries have banned or restricted DeepSeek?
| Country | Action reported by Reuters |
|---|---|
| Italy | Data protection authority blocked the chatbot (January 2025) |
| Australia | Banned on all government devices (February 2025) |
| Taiwan | Banned for government departments (February 2025) |
| South Korea | New downloads suspended in February, restored in late April 2025 |
| Germany | Asked Apple and Google to remove the app (June 2025) |
| Czech Republic | Banned in public administration (July 2025) |
| Netherlands | Civil servants told not to use it (July 2025) |
| France | Privacy regulator questioned the company |
| United States | Penalties weighed, lawmakers raised concerns |
| India | Finance Ministry staff advisory only (Dawn, citing Reuters) |
The pattern is consistent: most actions target government devices and official data, not private citizens. Italy and South Korea went furthest by blocking or pausing public access, and South Korea restored downloads after DeepSeek made changes.
Where does your data go when you use DeepSeek?
The answer depends on which route you use:
| How you use DeepSeek | Where data goes | Main risk |
|---|---|---|
| Official app or website | DeepSeek servers in China | Data location and policy |
| Official DeepSeek API | DeepSeek servers in China | Same as the app |
| Hosted by another provider | That provider's servers | Depends on the provider |
| Self-hosted open weights | Your own machine or cloud | Model safety only |
According to Tom's Hardware's analysis of DeepSeek's privacy policy, the company collects device model, operating system, IP address and system language, along with keystroke patterns and rhythms, your prompts, uploaded files and chat history. It stores this information on servers in China and may share it within its corporate group.
There has also been a real security incident. In January 2025, cloud security firm Wiz found a publicly accessible DeepSeek database containing more than a million lines of logs, including chat histories, secret keys and backend details, according to Axios. Wiz said the database allowed full control without authentication. DeepSeek secured it after being notified.
How safe is the DeepSeek model itself?
Several independent tests in 2025 examined how easily DeepSeek's models can be pushed into harmful behaviour:
- Cisco and the University of Pennsylvania tested 50 prompts from the HarmBench benchmark and reported that DeepSeek R1 failed to block any of them, a 100% attack success rate. The same test gave 86% for GPT-4o, 64% for Gemini 1.5 Pro, 36% for Claude 3.5 Sonnet and 26% for o1-preview. Cisco said the test cost less than $50.
- Palo Alto Networks' Unit 42 reported that jailbreaking techniques got DeepSeek to produce phishing emails, malware code and instructions for incendiary devices.
- Enkrypt AI reported that R1 was four times more likely than OpenAI's o1 to generate insecure code, and produced harmful content in 45% of its harmful-content tests.
The US government's evaluation
The US National Institute of Standards and Technology's Center for AI Standards and Innovation (CAISI) evaluated DeepSeek V3.1, R1 and R1-0528 against US models including GPT-5, GPT-5-mini, gpt-oss and Claude Opus 4 across 19 benchmarks, according to TechRepublic:
| Test | DeepSeek | US models |
|---|---|---|
| Agent hijacking (tried to steal login codes) | 37% | 4% |
| Complied with jailbreak requests | 95% to 100% | 5% to 12% |
| Cybench cyber tasks solved | 40% | 74% (GPT-5) |
| SWE-bench Verified coding | 55% | 63% to 67% |
| Echoed CCP narratives (Chinese prompts) | 12% | 2% to 3% |
CAISI also found that GPT-5-mini cost about 35% less on average than DeepSeek's best model to complete the same tasks. That contrasts with DeepSeek's reputation for low per-token prices, which we covered in DeepSeek price in India, and reflects how many tokens each model uses per task. CAISI is a US government body evaluating a Chinese competitor's models, so its framing should be read with that in mind, but its methods and benchmarks are published.
Should Indian users and developers use DeepSeek?
Based on the evidence, a reasonable approach is to match the route to the data:
- Casual questions with no personal data: the app carries limited risk, but assume anything you type is stored in China.
- Work data, customer data or code: avoid the official app and API. India's Digital Personal Data Protection rules allow the government to restrict cross-border transfers, as we explained in India's AI Data Centres in 2026.
- Developers building products: use DeepSeek through a provider that hosts the open weights outside China, or self-host. Our DeepSeek API guide for India covers the integration.
- Agents with tools or user-facing apps: add your own guardrails, because the model's built-in refusals are weak. Remove sensitive data before it reaches any model, as shown in our guide to anonymising data before an LLM.
- Government or regulated work: follow your organisation's policy. The Finance Ministry advisory covers ChatGPT as well as DeepSeek.
Frequently asked questions
Is DeepSeek banned in India?
No. DeepSeek is available to the public in India. The Finance Ministry advised its own staff in January 2025 to avoid AI tools including DeepSeek and ChatGPT on official devices, and the IT Minister said DeepSeek's models would be hosted on Indian servers.
Is DeepSeek safe to use?
It depends on the route. The official app and API store data on servers in China, and independent tests found its R1 model much easier to jailbreak than leading US models. Self-hosted open weights avoid the data issue but not the safety weaknesses.
Does DeepSeek send data to China?
Yes, if you use the official app, website or API. DeepSeek's privacy policy says personal information is stored on servers in China. Running the open-weight model yourself, or through a non-Chinese host, keeps data off DeepSeek's servers.
Which countries have banned DeepSeek?
Reuters reports that Italy blocked it, Australia and Taiwan banned it on government devices, South Korea paused downloads for about two months, the Czech Republic and the Netherlands restricted use in public administration, and Germany asked app stores to remove it.
Is DeepSeek safer if I run it locally?
Running the open-weight model locally means your prompts never reach DeepSeek's servers, which removes the data-location risk. It does not fix the model's weak refusals, so apps built on it still need their own guardrails.
How did DeepSeek perform in the Cisco security test?
Cisco reported a 100% attack success rate for DeepSeek R1 on 50 HarmBench prompts, meaning it blocked none. The same test gave 26% for o1-preview and 36% for Claude 3.5 Sonnet.
What did the US NIST evaluation find about DeepSeek?
NIST's CAISI found DeepSeek models lagged US models on most benchmarks, were far more susceptible to jailbreaking and agent hijacking, and echoed Chinese Communist Party narratives more often.
Limits of this review
- Security tests are snapshots. Most tested 2025 versions of R1 and V3.1, and newer releases may behave differently.
- Some testers sell security products. Cisco, Palo Alto Networks, Wiz and Enkrypt AI all have commercial interests in AI security.
- Policies change. Government positions and DeepSeek's privacy policy may have changed after the dates of the sources below.
- Not legal advice. Check your organisation's rules and the DPDP rules for your use case.
References
- Indian finance ministry asks employees to avoid AI tools like ChatGPT, DeepSeek, Dawn, citing Reuters, February 2025
- India lauds Chinese AI lab DeepSeek, plans to host its models on local servers, TechCrunch, 30 January 2025
- Which countries have banned or restricted DeepSeek, Insurance Journal, citing Reuters, 7 January 2026
- DeepSeek's privacy policy and the data it collects, Tom's Hardware
- DeepSeek database exposed, researchers find, Axios, 29 January 2025
- Wiz Research uncovers exposed DeepSeek database leaking sensitive information, Wiz, January 2025
- Evaluating security risk in DeepSeek and other frontier reasoning models, Cisco, 31 January 2025
- Recent jailbreaks demonstrate emerging threat to DeepSeek, Palo Alto Networks Unit 42, 2025
- DeepSeek R1 red teaming report, Enkrypt AI, 2025
- NIST evaluation finds DeepSeek models lag US AI and pose security risks, TechRepublic, 2025
- CAISI evaluation of DeepSeek AI models finds shortcomings and risks, NIST, September 2025