AI tools, connected apps and security layers

Learning edition: 5 October 2026. Introductory learning material.

Security tools need an owner who can configure, monitor and maintain them. Keep account controls, device protection, updates, logs, network boundaries, backups and response plans working together. For each control, record its purpose, owner, alert-response route and evidence that it works.

Look closely: Control the flow. Before connecting an AI tool, define the data it may read, the actions it may take and who checks the result. The dam is a metaphor; no filter guarantees safety.

Control Your Personal Data Stream by Elio Reichert, via UC Berkeley CLTC. CC BY 4.0. Unmodified image.

Understand what an AI tool can do

A chatbot may produce text. A connected AI assistant or agent may also read files, use apps and take actions. Start with a small, clearly defined task and limited access. Name who approves connections, checks activity and can stop the system. Keep important actions subject to meaningful review and enforce access limits in the connected systems.

When source material tries to give instructions

Prompt injection is an attempt to steer an AI through content it processes, such as a document, email or webpage. The content may try to make the tool disclose information or take an unrelated action. Treat outside content as material to analyze, not authority to expand the task.

A warning in the prompt or a filter is not a guarantee. Limit the data and tools the AI can reach; use system-enforced permission boundaries and review destinations and changes before approving actions. Reduce the possible harm if the model follows an unsafe instruction.

Practice with a fictional workflow

An assistant summarizes supplier brochures. One brochure includes a request to send an internal price list to an outside address. That request is part of the untrusted source, not the business owner’s instruction. Keep this workflow limited to reading the selected brochures and drafting a summary. It does not need access to the price list or permission to send email.

Your task: Write the approved purpose, data allowed, actions prohibited and responsible reviewer. Use fictional data for a trial. Decide how the owner will inspect activity, disconnect the tool and respond if it exceeds scope.

Check the output, too

AI can produce plausible errors without an attack. Check facts and sources before relying on an answer. Have a qualified person validate security changes; do not run generated commands just because they sound confident.

Complete the knowledge check below, then carry one relevant improvement into your 30-day plan.

Further reading: NCSC: Thinking carefully before adopting agentic AI (15 May 2026); NCSC: Prompt injection is not SQL injection (8 December 2025); ASD ACSC, NCSC NZ and COSBOA: Artificial intelligence for small business (2025 edition).