Secure DeepSeek in a customer onboarding assistant

Question

Your company wants to use DeepSeek models (DeepSeek-V3 for chat, DeepSeek-R1 for harder reasoning) in a customer onboarding assistant that collects identity details, reads uploaded documents, and can call tools to create accounts and schedule verification calls. Security review flagged prompt injection, data leakage, and unauthorized tool access. As a staff engineer, how do you build the threat model and the mitigations, including the choice between the hosted DeepSeek API and self-hosting the open weights?

Accepted Answer

I start with what the assistant can touch: customer PII, uploaded documents, and tools that create accounts, then list how each could be abused. Because onboarding data is regulated PII, I would not send it to the hosted DeepSeek API unless legal approves its data-processing and residency terms; the usual choice is the open weights self-hosted in our own environment or on an approved cloud platform. The core controls are the same as for any model: tools authorized in code rather than by the model, uploaded documents treated as untrusted input, PII minimized and redacted, and R1's reasoning output never shown to users or logged raw.

Secure DeepSeek in a customer onboarding assistant

The interview question

Deep explanation