← All insights

Secure engineering

Five security controls every AI agency should consider before deployment

The most important security decision for an AI agent is often not which model you use. It's what authority you give the system, what it can see and what it can change.

Here are five practical controls to consider when building tool-connected AI solutions for customers.

1. Know where the trust boundaries are

Map the full path from input to action. User messages, retrieved pages, emails, uploaded documents and tool responses can all carry untrusted material. Explicitly decide where lower-trust information enters the agent's context and where it might influence a high-trust decision.

2. Give tools the smallest useful permissions

If an agent only needs to read a subset of customer records, don't give it broad edit rights across the database. Limit permissions by resource, operation, customer tenancy and environment where the platform allows it. Tie the scope to the human or service identity responsible for the action.

3. Add human approval where the impact justifies it

Not every reply needs an approval button. But changing a bank detail, deleting a customer account or sending a high-stakes external message should trigger a deliberate decision. The approval must occur at the action boundary, not just as a line in the prompt.

4. Keep an audit trail of meaningful decisions and actions

You need enough information to reconstruct a sequence: who initiated the workflow, which tool was called, what it attempted and what approval existed. Protect sensitive data in your logs, and set retention rules appropriate to the business and legal context.

5. Test the agent as a system, not a single prompt

Combine familiar app and API security controls with adversarial tests for prompt injection, tool misuse, cross-tenant access and sensitive data exposure. Retest after changing connected tools, permissions, retrieval sources or prompts.

The bigger point

These controls are most useful when they're discussed during design, integrated into the engineering workflow and verified ahead of deployment. They are not a guarantee of security, but they reduce the opportunity for avoidable failures.

Next step

Building an AI application for a customer?

Explore Secure AI Engineering