Is DeepSeek Safe? The Australian Ban and Privacy

Is DeepSeek Safe for Client Data?

DeepSeek’s hosted service is a poor choice for confidential legal material because the firm sends that material to a provider processing data in China. DeepSeek’s terms place disputes under mainland Chinese law. Its privacy policy also permits disclosures in response to applicable law or government requests.

For a lawyer considering a document translation, the question is what leaves the firm when the document enters the chat. A witness account or draft advice can reveal confidential information even after someone removes the client’s name. Use public or fabricated material for experimentation and keep client documents out of the hosted service.

The country of storage affects which authorities can seek access and which legal remedies are practical. Our data sovereignty guide explains that distinction for legal teams.

Is DeepSeek Banned in Australia?

Yes, DeepSeek is restricted on covered Australian Government systems and devices under PSPF Direction 001-2025. The Department of Home Affairs still lists the direction among its protective security directions.

The DeepSeek direction binds non-corporate Commonwealth entities under the Public Governance, Performance and Accountability Act 2013. These include Commonwealth departments. It requires them to prevent access, use and installation, and to remove existing instances. Covered devices include authorised non-government devices used within that government environment.

The direction permits a narrowly limited exemption process for national security and regulatory functions. An ordinary request to translate a client document does not fit that exception.

Private firms and members of the public are outside the direction’s general scope. DeepSeek remains a public online service, although a workplace can block access. A working website is not permission to use it for a matter. Government lawyers must also follow their own agency’s rules, including any separate state or territory restrictions.

What the DeepSeek Privacy Policy Says

DeepSeek’s privacy policy says it collects user inputs, including uploaded files and chat history. It also collects account details and device or network information, such as IP addresses. The policy places direct collection, processing and storage of personal data in the People’s Republic of China.

Its stated uses include training and improving models. Retention depends on the data and purpose, with some information kept while an account exists and further retention allowed for legal or business reasons. The policy provides rights to request deletion and opt out of training, subject to applicable conditions. Neither option changes the country where an upload is processed.

Security researchers have also found a concrete failure in the hosted infrastructure. In 2025, Wiz reported an exposed DeepSeek database containing more than a million log lines, including chat history and authentication secrets. Wiz said DeepSeek secured the exposure after disclosure. That finding describes a remedied incident, not proof that the same database remains open.

For legal teams, the incident shows why a provider’s storage policy and its security controls need separate assessment. A promise about data use does not prevent an access-control failure.

Running DeepSeek Models Locally

Yes, an open-weight DeepSeek model can run on a firm’s own hardware without sending prompts to DeepSeek when all processing stays local. DeepSeek publishes model weights and local-running instructions. That is a different arrangement from using its website or hosted API.

The firm needs suitable hardware and someone responsible for the software, access controls and stored prompts. Connected search, cloud fallback or external logging can still send information elsewhere. Assess the complete application, including its network connections, before treating it as local.

The government direction excludes qualifying locally deployed open-source models where the codebase is available for inspection and appropriate mitigations exist. A download alone does not establish that an agency meets those conditions.

Our private AI guide explains the hosting choices. For the practical requirements of running a local LLM, assess the firm’s hardware and support capacity before selecting a model for client work.