What is zero trust?
Zero trust is a security model that grants no access on the basis of network location. A request from inside the head office gets the same scrutiny as one from a hotel lobby. Every user, device, and service proves its identity, and each request is checked against policy. Access is granted to one resource at a time, following the principle of least privilege. The usual shorthand is “never trust, always verify.” The standard reference is Special Publication 800-207 from the U.S. National Institute of Standards and Technology (NIST).
Zero trust replaces the perimeter model, in which a firewall and a virtual private network (VPN) marked everything inside as trusted and an intruder could move freely once in. In a zero-trust design, a policy engine weighs each request against the user’s identity and role, the device’s health, the resource’s sensitivity, and signals such as an unusual location. Networks are divided into small segments, services authenticate to each other with short-lived credentials, and every decision is logged.
Zero trust and private AI
AI systems add new resources to protect: inference endpoints, vector databases, document indexes, and the tools that agents call. Running models on infrastructure the company controls removes a third party from the data path, but it does not make those components trustworthy by default. Many open-source inference servers answer anyone who can reach them unless authentication is configured. Agents need identities of their own and should act with the permissions of the user they serve. A zero-trust AI agent, for example, receives short-lived credentials for one task and has each tool call checked against policy.
Zero trust in practice
A manufacturer of hydraulic components runs a maintenance assistant on servers in its own data center. Technicians reach it only from company-managed tablets, after single sign-on with multi-factor authentication. The inference server sits in its own network segment and accepts requests only from the assistant application. The assistant reads work orders through the maintenance system’s interface but has no network route to the machine controllers on the plant floor. A stolen password used from an unmanaged laptop gets nothing, and a manipulated assistant cannot reach production equipment.
What to watch for
- Zero trust is a design approach, so ask each vendor which parts of it the product covers and which it leaves to other systems.
- Include service-to-service traffic, such as ingestion jobs that write to a vector database, which perimeter-era designs often leave open.
- Plan for legacy systems that cannot handle modern authentication, usually by placing an identity-aware proxy in front of them.
- Roll out in stages, starting with the systems that hold the most sensitive data.