AI Security Checklist for Australian Businesses

An AI security checklist for an Australian business comes down to seven checks: know which AI tools your staff actually use, set a written policy on what data may go into them, vet each vendor, control accounts and access, keep a human reviewing outputs, meet your privacy obligations, and have an incident plan ready. This page walks through each one in order, with the Australian guidance behind it, so you can work down the list this week.

TL;DR

• You cannot secure AI tools you cannot see. Start with an honest inventory of what staff use, including free and personal accounts.
• The Australian Cyber Security Centre (ACSC) says to treat AI like any other online service: useful, but not automatically private or secure.
• Privacy law still applies. Personal information you put into an AI tool, and personal information it generates, falls under the Privacy Act 1988.
• Write a one-page AI use policy, vet vendors, review outputs, and prepare an incident plan before you need one.
• Treat the checklist as a quarterly routine, not a one-off tick-box.

Table of Contents

• Why a checklist, and why now
• The checklist at a glance
• 1. Find out what AI your people already use
• 2. Decide what data may go into AI
• 3. Write a one-page AI use policy
• 4. Vet each AI vendor
• 5. Lock down accounts and access
• 6. Keep a human in the loop
• 7. Meet your privacy obligations
• 8. Train staff and make reporting easy
• 9. Prepare an AI incident plan
• 10. Review every quarter
• Our perspective: visibility comes before controls
• How Calibre Forum fits in
• FAQ
• Sources

Why a checklist, and why now

Most AI risk in a small or mid-sized business does not come from someone building a model. It comes from an employee pasting a client contract, a spreadsheet of customer details or a draft strategy into a chatbot to save twenty minutes. The ACSC’s guidance for small business on AI lists the main risks as data leaks and privacy breaches, unreliable outputs and supply chain vulnerabilities. None of these need a sophisticated attacker.

The common failure is a “tool first, security second” approach: the tool spreads through the team, then someone asks what it does with the data. The checklist below reverses that order. It is written for owners, managers and whoever looks after IT, and it does not assume you have a security team.

The checklist at a glance

1. Find out what AI your people already use

You cannot apply any control to a tool you do not know about. Staff adopt chatbots, meeting transcribers, browser extensions and image generators on their own, often on personal accounts. This is usually well meant, which is why punishing it backfires: people stop telling you.

To build the inventory:

• Ask every team, not just IT, what they use and for which tasks.
• Check browser extensions, and AI features switched on inside software you already pay for.
• Record the tool, the plan (free, paid, business), who uses it and what data goes in.
• Make answers anonymous if you can. People answer more honestly when no one is being audited.

The last point is why we built our assessment around an anonymous staff survey. More on that below, and in our guide to finding shadow AI in your business.

2. Decide what data may go into AI

You need a simple answer to “what can I paste into this?” Most businesses do well with three tiers:

• Open: published marketing copy, public information, generic questions.
• Internal: non-public but low-harm material, such as meeting agendas. Only in approved business-grade tools.
• Restricted: customer or patient details, health, financial or HR records, credentials, contracts and anything commercially sensitive. Never in public AI tools.

The OAIC’s position is clear on the top tier. As a matter of best practice, it recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. Our practical guide to AI data classification shows how to build these tiers without a consultant.

3. Write a one-page AI use policy

The ACSC’s key actions begin with establishing policies for how AI tools can be used and what information can be shared with them. A policy that works is short enough to be read. Cover:

1. Which tools are approved, and how to request a new one.
2. What data is banned from AI tools (your restricted tier).
3. That personal accounts are not to be used for work data.
4. That a person reviews anything AI-generated before it goes to a client or informs a significant decision.
5. Who to tell when something goes wrong, and that reporting will not be punished.

Put it in onboarding and repeat it at a team meeting. A policy no one remembers is not a control.

4. Vet each AI vendor

Many owners assume that data entered into an AI tool is automatically protected by the vendor. It is not. The ACSC’s small business handbook says to consider whether the provider is reputable, whether you really need to share the information requested, and how and where your data will be stored, used or shared.

A workable vendor review asks:

• Does the vendor train its models on your inputs, and can you switch that off?
• Where is data stored, and for how long?
• Can you delete data and close the account fully?
• Does it publish security certifications such as ISO 27001, or align to a recognised framework?
• Does it offer single sign-on, multi-factor authentication and admin controls on the business tier?
• What happens to your data if the vendor is sold or shuts down?

Free and consumer plans often differ from business plans on exactly these points. We have condensed the process into a one-page AI vendor risk checklist you can hand to whoever approves purchases.

If your marketing team experiments with generative video, music and image apps, such as those used for AI short film creation, run them through the same questions. Uploaded footage, voices and brand assets are business data too.

5. Lock down accounts and access

Security guidance for AI systems leans on a principle familiar from the rest of IT: least privilege. The ACSC’s AI data security publication calls for strong access controls and encryption to protect training and operational data, and for validating and monitoring data inputs.

For a business using off-the-shelf AI, that translates to:

• Business accounts with multi-factor authentication, not shared logins.
• Only the people who need a tool have it, with admin rights held by one or two people.
• Connectors and integrations (to email, drive, CRM) reviewed before switching on, since each one widens what the tool can read.
• Chat history and data retention settings configured deliberately.
• API keys stored in a password manager or secrets vault, never in documents or chat.
• Accounts removed when staff leave.

6. Keep a human in the loop

AI output can be wrong, confident and persuasive at once. The ACSC’s guidance asks businesses to verify AI-generated outputs and maintain human oversight for important decisions. Practical rules:

• AI drafts and summarises. A named person checks facts, figures, names and tone before anything is sent externally.
• High-stakes decisions (legal, medical, financial, hiring) are never made by AI alone.
• Reviewers check for invented references, which is a known failure of language models.
• Anything customer-facing is labelled as AI-assisted where appropriate.

The OWASP Top 10 for LLM Applications 2025 names prompt injection and sensitive information disclosure among the leading risks, and a crosswalk that maps them to the NIST AI Risk Management Framework shows how to organise governance under its four functions: govern, map, measure and manage. You do not need to adopt either in full, but they are a useful reference if a larger client asks how you manage AI risk.

7. Meet your privacy obligations

This is the area where Australian businesses most often misjudge their position. The OAIC states that privacy obligations apply to any personal information input into an AI system, and to output that contains personal information. That includes outputs that are inaccurate. If a tool generates a statement about an identifiable person, it can still be personal information.

Checks to complete:

• Map where personal information could enter AI tools, and block it where you can.
• Update your privacy policy so it states how and why AI is used, particularly where information is used for a secondary purpose.
• Tell customers when they are dealing with an AI-enabled tool, such as a chatbot.
• Complete a privacy impact assessment before rolling out any AI feature that handles personal information.

This article is general information, not legal advice. If you handle health, financial or children’s data, or you are unsure whether a use of AI is permitted under the Australian Privacy Principles, speak to a privacy lawyer or your industry regulator before you proceed.

What about voluntary standards and future rules?

Australia’s Voluntary AI Safety Standard, published by the National AI Centre on 5 September 2024, sets out ten guardrails and does not create new legal duties. It has since been pointed towards the newer Guidance for AI Adoption. Rules for high-risk AI have been debated, so check current government announcements rather than relying on any single summary. Adopting sensible controls now is cheaper than retrofitting them later.

8. Train staff and make reporting easy

Policies fail when staff do not know them. Training should be short and concrete: show real examples of what not to paste, how to spot an AI-written phishing email, and what a suspicious AI prompt or file looks like. The ACSC’s guidance includes training staff on secure and responsible AI use as a core action. Our overview of what staff need to know about AI security is written for non-technical teams.

Reporting matters as much as training. If someone pastes a client list into a chatbot by mistake, you want to hear within minutes, not months. Make the report channel one email address or one chat message, and make clear that nobody is disciplined for honest, prompt reports. Our piece on how AI data leakage happens and how to stop it covers the usual routes.

9. Prepare an AI incident plan

The Notifiable Data Breaches scheme trips many businesses up. A guide on AI data breaches and the NDB scheme explains the common misreading: you have 30 days to assess whether a breach is eligible, not 30 days to notify. If it is eligible, notification to the OAIC must happen as soon as practicable.

Your plan should answer, on one page:

• Who owns it: a named person or role, with a backup.
• First hour: how to revoke API keys, disconnect integrations, change passwords and, where the vendor allows, purge data from a workspace.
• Evidence: what to record (tool, account, data involved, time, who).
• Assessment: who decides whether the incident is likely to cause serious harm.
• Contacts: the vendor’s security contact, your insurer, your legal adviser.

Rehearse it once with a made-up scenario. An hour of rehearsal is cheap compared with a real incident.

10. Review every quarter

The most common mistake is treating AI security as a checkbox. The ACSC’s AI data security guidance describes risk across the full AI lifecycle, and tools, plans and features change constantly. A vendor that was safe in January may have changed its terms by April.

Every quarter, repeat the inventory, check for new tools, confirm that policy and training still match reality, and re-run vendor checks on anything that changed.

Our perspective: visibility comes before controls

We see the same pattern in Australian and New Zealand businesses. Leadership believes the team uses one or two approved AI tools. Staff, when asked privately, use six or seven, and a few of them touch customer information. Nobody is acting in bad faith. The tools are simply easier than the official process.

That is why we do not start with software on laptops or monitoring of individuals. We start with a conversation, run as an anonymous survey of every staff member, because people tell the truth when they are not worried about being identified. The result is a map of real usage that no policy document can give you, and a place to begin the checklist above with facts instead of guesses.

Our view is that a checklist is only as good as the inventory beneath it. Do that first, keep it current, and most of the other controls become a matter of sequencing.

— Anthony

How Calibre Forum fits in

Our AI readiness assessment surveys your staff anonymously about the AI tools they use and how they use them, then combines the answers into a scored report with a tailored 90-day action plan covering operational and security improvements. We do not install software or collect individual data, and we do not make compliance or legal claims, so it works alongside your lawyers and IT provider, not instead of them.

The Report is A$1,990 one-off. If you want to repeat the exercise as the checklist suggests, the Membership is A$2,388 per year, with quarterly assessments and reports. You can see what the output looks like in our sample AI readiness report, or read how the process runs from setup to report.

If you run a medical clinic, an aged care or disability provider, a childcare centre, a training organisation or a managed service provider, we also have pages written for your sector, such as our assessment for medical clinics and for aged care and NDIS providers.

FAQ

Do small businesses really need an AI security checklist?

Yes. The ACSC wrote its AI guidance specifically for small business owners and staff who use cloud-based AI tools. Small teams often have fewer approvals in place, so one employee’s choice of tool can expose customer data. A short checklist is proportionate to that risk.

Is it safe to put customer information into ChatGPT or similar tools?

The OAIC recommends, as best practice, that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. If you need AI for tasks involving personal data, use a business-grade product with terms you have reviewed, and get privacy advice first.

Who should own AI security in the business?

Name one accountable person, usually the owner or operations manager, and involve IT, HR and team leads. The ACSC and OWASP guidance both point to AI security being a cross-functional job, not only an IT task.

How often should we review our AI tools and policies?

Quarterly is a sensible rhythm for most small and mid-sized businesses, since vendors change terms and staff adopt new tools constantly. Review sooner if you roll out a new AI feature or have an incident.

Sources

• Artificial intelligence for small business (ACSC)
• Using AI securely in your business (ACSC)
• AI data security (ACSC)
• Guidance on privacy and the use of commercially available AI products (OAIC)
• OWASP LLM Top 10 and NIST AI RMF crosswalk
• AI Data Breaches and the NDB Scheme (Need to Know AI)
• Voluntary AI Safety Standard (10 Guardrails)

Recommended

• AI Security Awareness: What Your Staff Need to Know
• AI Data Classification: A Practical Guide for Australian Businesses
• AI Security Assessment for Business Leaders: Find Shadow AI

General information only. Articles on this blog are written with the help of AI writing tools and published by Calibre Forum. They are not legal, insurance, tax or financial advice, so check anything important with a qualified adviser.