AI Security Awareness: What Your Staff Need to Know

AI security awareness means your people know what they can and cannot put into an AI tool, can spot AI-powered scams, and know who answers for the output. It is not a one-off video or a policy PDF. It starts with seeing how your staff already use AI, because you cannot train for habits you have not seen.

This guide is for Australian business owners, managers and IT leads. It covers the risks that matter, what Australian regulators say, what good training contains, and a 90-day plan you can start this month.

Calibreforumcalibreforum.auSee How Your Team Uses AICalibre Forum’s structured staff survey reveals current AI use and produces a scored report with a clear 90-day action plan.Explore the AI readiness assessment

TL;DR

Most AI security incidents in small businesses start with ordinary staff pasting client or personal information into a public tool, not with sophisticated attacks.
The OAIC advises that organizations covered by the Privacy Act avoid entering personal information, especially sensitive information, into publicly available generative AI tools.
A policy alone is not enough. Pair it with practical training, technical guardrails and named checkpoints where a person reviews AI output.
Start by measuring real usage with an anonymous staff survey, then train against what you find.
A 90-day cycle (see, set rules, train, re-measure) is realistic for a small or mid-sized team.

What AI security awareness actually covers
Why awareness is the weak point
What Australian regulators say
What good AI security training contains
Why start by listening to staff
A 90-day plan for a small or mid-sized business
Our perspective
Getting this right in your industry
See your own picture first
FAQ
Sources

What AI security awareness actually covers

Search results for this topic mix two different things. One is AI used by security teams to detect threats. The other, which this article is about, is the human side of AI at work: how ordinary employees use chatbots, copilots and AI features inside everyday software without creating a security or privacy problem.

In practice it covers four areas:

Data handling. What information may go into which tool, under which plan.
Threat recognition. Deepfakes, cloned voices and AI-written phishing that look more convincing than the old, clumsy kind.
Output accountability. AI makes things up. Someone has to check before it reaches a client.
Tool visibility. Knowing which tools are actually in use, including the ones nobody approved.

The National AI Centre treats this as part of responsible use, not an IT afterthought. Its essential AI practices are the Australian Government’s guidance for putting governance, oversight and transparency in place.

Why awareness is the weak point

The gap between what leaders assume and what staff do is large. A 2025 survey behind the National Cybersecurity Alliance’s AI Fools campaign found that 58% of users report receiving no security or privacy training on AI, and 43% admit sharing sensitive work information with AI without their employer’s knowledge. That is a global survey, not an Australian one, but there is no reason to assume your team is the exception until you have checked.

People rarely do this out of carelessness. They are trying to finish a task faster: summarize meeting notes, tidy a client email, draft a tender response. The tool is open in a browser tab, free, and nobody told them not to.

Why awareness is the weak point — overview diagram

Shadow AI is a visibility problem first

If you ask a manager which AI tools the team uses, you get the approved list. If you ask the team, anonymously, you get the real list. The difference is where your risk sits. You cannot write a sensible rule for a tool you do not know exists, and staff who fear blame will not volunteer it.

Anonymous staff reports reveal unapproved AI tools

The subscription tier matters

The Australian Cyber Security Centre’s guidance on AI for small business stresses that small businesses need to take proactive steps to protect data, customer privacy and business systems. One practical point leaders often miss is the vendor’s data terms: whether inputs are retained or used to train models can differ between free, paid and enterprise tiers, and opt-outs are not always the default. Someone needs to read those terms for each tool, and staff need to know which tier they are on.

What Australian regulators say

You do not need to become a lawyer, but you should know the reference points. This is general information, not legal advice. If you handle health, financial or other sensitive records, speak to your legal adviser about your specific obligations.

Privacy: the OAIC position

The OAIC’s guidance on commercially available AI products (October 2024) says privacy obligations apply to personal information put into an AI system, and to output that contains personal information. Its later blog on GenAI tools in the workplace (December 2025) restates the advice that organizations subject to the Privacy Act refrain from entering personal information, particularly sensitive information, into publicly available tools.

That same OAIC blog makes a second point worth taking seriously: a written acceptable use policy on its own does not cover you, because people make mistakes. It points toward layered measures, such as technical controls and continuous training, working together.

Cyber security: the ACSC position

The ASD’s ACSC published Engaging with Artificial Intelligence, guidance on using AI systems securely, co-sealed with international partners. Its small business page is the more practical read for most teams and is where to start.

Governance: it is a leadership question

The government’s page on AI and Australian law is a good starting point for directors. The practical lesson: AI risk cannot simply be handed to IT. Leaders need to understand what the AI is used for and how it works well enough to oversee it.

What good AI security training contains

Generic online courses cover “what is AI” and ethics. Useful, but they do not change what Priya in accounts does on Tuesday afternoon. Training that works is built around your actual tools and tasks. Here is what we would put in it.

1. A clear traffic-light rule for data

Staff need a simple answer to “can I paste this?”. A three-tier rule is easy to remember:

Category	Examples	Rule
Red	Client names, health or financial records, staff HR matters, credentials, unreleased commercial terms	Never into a public tool
Amber	Internal procedures, anonymized drafts, de-identified case notes	Only in approved tools, after removing identifying details
Green	Public information, generic questions, your own published content	Fine in approved tools

Your own categories will differ by industry. A clinic’s red list is longer than a landscaping firm’s.

2. Scrubbing before prompting

The OAIC blog describes a habit worth teaching: treat what you type into a public AI tool as if you were posting it on a public noticeboard, and remove personal details or confidential markers first. Practise it with real examples from your business. Take a genuine client email, strip the identifiers, then ask the question.

3. Recognizing AI-assisted scams

The National Cybersecurity Alliance notes that scammers use AI to craft convincing deepfakes, impersonate trusted contacts and trick people into sharing sensitive information. The practical defense is procedural: a call-back rule for payment changes, a second person for urgent transfers, and a code word for sensitive requests that come by voice or video.

4. Human checkpoints for AI output

The National AI Centre’s guidance on redesigning workflows says that for each task you improve, you should be clear about what AI does, what people do, and how outputs are checked. Turn that into named checkpoints. For example: any AI-drafted text going to a client is read in full by the sender; any figure is traced to its source; any odd or invented answer is reported to a named person.

5. Knowing what to do after a mistake

Someone will paste the wrong thing eventually. If staff fear punishment they will hide it, and you lose the time when containment is possible. Say plainly who to tell, that reporting quickly is expected, and what happens next.

Why start by listening to staff

Most businesses begin with a policy and a training session. We suggest reversing the order. Before you tell people what to do, find out what they already do.

An anonymous survey of every staff member, not a sample of managers, tends to surface three things:

Tools you did not know were in use, and for which tasks.
Where staff are unsure of the rules, which tells you what training to write first.
Where AI is genuinely saving time, which you want to protect rather than ban.

Anonymity matters because honest answers matter. If people think an answer could be traced back to them, they give the answer they think you want.

A 90-day plan for a small or mid-sized business

This is a framework we use for thinking about the work. Adjust it to your size and industry.

Days 1 to 30: see

Run an anonymous survey of all staff on which AI tools they use, for what, and how confident they feel.
List every tool mentioned, and check each vendor’s data terms for your subscription tier.
Identify which tasks touch personal or client information.

Days 31 to 60: set rules and guardrails

Publish the traffic-light rule and the approved tool list. Keep both to one page.
Provide an approved option for the tasks people clearly want AI for, so the rule is not just “no”.
Add technical controls where your IT setup allows them, as the OAIC suggests alongside training.
Define the human-in-the-loop checkpoints for client-facing work.

Days 61 to 90: train and re-measure

Run short, role-specific sessions using real, scrubbed examples from your own business.
Brief directors and senior managers, since oversight sits with them.
Survey again with the same questions. Compare. Fix what did not move.

Then repeat on a regular rhythm. Tools change quickly, and so do staff habits.

Our perspective

We built Calibre Forum around a simple observation: most leaders lack a picture of how AI is actually used inside their business. They have a policy, perhaps a training module, and a feeling that things are probably fine. What they do not have is evidence.

So our assessment starts with the people doing the work. Every staff member completes a structured, anonymous survey. We combine the responses into a scored report that maps current capabilities, flags risks and lays out a 90-day action plan. We do not install software on your devices and we do not collect individual-level data. We also do not make compliance or legal claims: the report gives you visibility and next steps, and your legal and IT advisers decide what your obligations require.

We think awareness training works best when it follows measurement. A session aimed at the wrong problem wastes everyone’s afternoon, and a banned tool that staff keep using in secret is worse than an approved one they use openly.

— Anthony

Getting this right in your industry

The risks look different depending on who your clients are. A few examples of where the specifics change:

Health and care. Records are sensitive by nature. See how we approach medical clinics and aged care and NDIS providers.
Professional services. Client confidentiality is the product. Relevant pages cover accounting firms and law firms.
IT providers. You are managing AI risk for your clients as well as yourself: managed service providers.

See your own picture first

If you want the survey-first approach without building it yourself, the Calibre Forum assessment is a one-off Report at A$1,990, or a Membership at A$2,388 per year for recurring assessments and reports. You can look at the sample report first, or read what is in the readiness report.

FAQ

Is a staff AI policy enough?

No. The OAIC’s workplace blog argues that internal acceptable use policies alone are not sufficient, because human error is inevitable. Combine the policy with training, technical controls where possible, and checkpoints for reviewing AI output.

Can employees use free AI tools with client information?

For organizations covered by the Privacy Act, the OAIC advises against entering personal information, particularly sensitive information, into publicly available generative AI tools. When in doubt, use de-identified examples or an approved tool, and check your obligations with a legal adviser.

How often should we run AI security training?

An annual session is rarely enough, because tools and habits change quickly. A sensible rhythm is an initial round of training, then short refreshers and a repeat measurement every quarter or so, adjusted to your risk level.

Who is responsible for AI risk in a business?

Leadership. The government’s guidance on AI and Australian law points out that directors need to understand how AI is used rather than leaving the matter to IT. IT supports the controls, but accountability sits higher.