Zero trust for AI systems

AI adds new things to protect: training data, models, prompts and, increasingly, the actions systems can take. Zero-trust principles apply to all of them.

A glowing teal LED panel
Photo: Adi Goldstein on Unsplash

Zero trust is a simple idea with a demanding implementation: never assume that a user, device or service is safe just because it is inside the network. Verify every request, grant the least privilege needed and record everything.

As AI systems move from answering questions to taking actions, these principles matter more than ever.

What needs protecting

  • Data: the records and streams models learn from and act on.
  • Models: their weights, versions and the pipeline that updates them.
  • Inputs: prompts, documents and sensor feeds that could be manipulated.
  • Actions: the alerts, workflows and integrations a system can trigger.
Verify every requestLeast-privilege accessSigned models & dataContinuous monitoring
Zero trust applied to AI: never assume, always verify.

Principles we apply

Least privilege for machines, too

An AI component should only be able to read the data and trigger the actions it genuinely needs. An assistant that summarises documents has no business sending messages or changing records.

Verify every model you run

Models should be versioned, signed and deployable only through a controlled process, with the ability to roll back. An unverified model is an unverified piece of code.

Treat inputs as untrusted

Documents, messages and even camera feeds can be crafted to mislead a model. Validate inputs, separate instructions from content and monitor for unusual behaviour.

Approve the actions that matter

High-impact actions should require human approval, with the evidence attached and the decision logged.

Security for AI is not a new discipline. It is the old discipline, applied honestly to new things.

Certifications are a floor, not a ceiling

Our information security management system is certified to ISO/IEC 27001:2022. Standards like this provide a solid baseline. The work of applying them to models, inputs and autonomous actions is where institutions now need to focus.

Want to explore how this applies to your organisation? We share our approach with qualified organisations in a private briefing.