Skip to content
Cyber Unboxed
AI Security

Securing AI Systems: The Attack Surface Nobody Draws

An AI feature is not just a model. It is data, prompts, tools, permissions and people, and each is something to secure.

2 min readIntermediate Aug 17, 2026

Explain Like I'm Not a Hacker

Giving an AI access to your things is like hiring a very helpful assistant: check what keys they hold and who is allowed to give them instructions.

The 30-second explanation

An AI assistant connected to your files, email and tools is like an employee with a lot of access. The question is what it can reach, who can talk to it, and what happens when it is tricked.

How it works

  1. 1

    1. Map

    Draw data, model, tools and outputs.

  2. 2

    2. Limit

    Give the system the least access it needs.

  3. 3

    3. Validate

    Check inputs and outputs, not just prompts.

  4. 4

    4. Monitor

    Log tool calls and review odd behaviour.

Start by drawing the whole system: where data comes from, what the model can read, what tools it can call, and where its output goes. Then apply familiar controls: least-privilege access for tools, validation of outputs before they are used, secrets kept out of prompts, logging of tool calls, and review of third-party models and plugins. Because models can be steered by the content they read, treat everything they ingest as untrusted input.

Real-world example

A support chatbot is connected to a customer database so it can answer account questions. A review finds it can read every customer's record. The team limits it to the current customer's data and logs each lookup.

How to spot it

  • Broad tool access

    An assistant that can reach more than its task needs.

  • Secrets in prompts

    Keys or passwords included in instructions or context.

  • Unvalidated outputs

    Model output used directly in commands or queries.

  • No logging

    No record of what the assistant did or was asked.

What to do

  1. 1Draw the data and tool flow for every AI feature before launch.
  2. 2Apply least privilege to every tool and data source the model can reach.
  3. 3Log requests and tool calls, and review them regularly.

Stay curious. Stay safer.

This is one piece of a bigger picture. Explore more real-world examples, concepts and tips to build your cybersecurity awareness.

Explore More

Keep reading