AI threats

A familiar voice is no longer proof. Check the request, not the caller.

Attackers use the same AI tools your team does: to write perfect phishing emails, clone a director's voice, fake a video call and trick AI assistants. Here are the six patterns we see, what they look like, the controls that work, and a check to run before you act on a request.

Agentic AI guidance from ASD and Five Eyes partners, checked 31 October 2026.

Book an AI threat briefing Check a request 07 3114 2808

Six attack patterns
Controls that work
Verify-first check

Reviewed by Blake Bath, General Manager, on 31 October 2026.

A woman in an office at dusk holds a desk phone to her ear while a video call plays on her laptop, checking a request on a second channel. — **Call back, every time**Payments, bank changes and resets get checked on a channel you start yourself.

What we see

Four AI attacks that target people.

Each one is built to get past someone in a hurry.

Deepfake video calls

A 'director' on a short video call asks for an urgent, confidential payment.

Call back on a known number
Two approvers
Staff briefed

Verify-first check

Got a request that feels off? Check it before you act.

Four questions about the request. You get the exact verification step for that kind of request, ready to copy.

Answer four questions before you act. Nothing you choose leaves this page.

What are they asking you to do?

How did the request arrive?

Is it urgent or confidential?

Does it seem to come from a manager or director?

Verify first

3warning signs

Call the requester back on the number in your own records, then follow the normal two-person approval.
Use an email to their known address and a call back you start yourself, not a reply to this request.
If anything does not match, stop and tell IT. A real request will survive a check.

Voices and faces can now be faked from a short sample of public audio or video.
Urgency and secrecy are the pressure the scam relies on.
Requests that appear to come from leaders are the most common target.

Try the incident simulator

The full list

Six AI-driven threats and the controls that help.

Threat	What it looks like	Controls that help
Deepfake CEO or CFO fraud	A short video or call from a 'director' asking for an urgent, confidential payment	Call back on a known number; two-person approval for payments
AI-written email fraud	Polished emails that mention real projects and people, asking to update payment details	DMARC at reject, mailbox rule alerts, phone verification
Voice cloning	A phone call in a manager's voice asking for a password reset or payment	No resets or payments on voice alone; helpdesk verification script
Prompt injection	An AI assistant behaving oddly after reading an external document or web page	Trusted content only, business AI tools with logging, sensitivity labels
Faster exploitation	Known flaws in internet-facing systems used as the way in. Verizon's 2026 report puts this as the top first step in breaches, at 31%	Patch internet-facing systems first and fast, watch your attack surface
AI agents doing too much	An agent taking actions outside its job after reading a malicious instruction	Least privilege, human approval, logging, start small

Official guidance

AI agents need tight limits.

On 1 May 2026, CISA, ASD's ACSC and Five Eyes partners published joint guidance on securing agentic AI. The practical points for a small business:

Least privilegeGive each agent only the access its job needs.
Human approvalA person approves high-impact or sensitive actions.
Monitoring and logsRecord what agents do so it can be reviewed and reversed.
Start smallLow-risk tasks first, more autonomy only as confidence grows.

Fifteen minutes

Three messages for a fifteen-minute staff brief.

Good writing is no longer a warning signJudge a request by what it asks for. Verification is the safety net.
A voice or video is not authorityPayments, bank changes, resets and MFA changes are checked outside the call, every time.
AI tools can be trickedOnly give them content you trust. Report odd behaviour.

Repeat it every quarter. It takes fifteen minutes and covers most of the risk.

Our view

The defence is mostly process, not more AI.

AI makes attacks more convincing, but the money still moves through the same few actions: a payment, a bank detail change, a password reset. A written rule that these are always verified on a separate channel beats any detection product.

On the technical side, phishing-resistant MFA, DMARC at reject and fast patching of internet-facing systems blunt most of what AI adds. Then use business AI tools with logging, and keep agents on a short leash.

The order

Getting ready for AI-driven attacks in five steps.

Step 1Write the verification rule
Payments, bank changes, resets and MFA always checked.
Step 2Brief the team
Fifteen minutes, three messages, every quarter.
Step 3Harden email and sign-in
DMARC at reject, phishing-resistant MFA.
Step 4Patch fast
Internet-facing systems in days.
Step 5Govern AI use
Approved tools, logging, limits on agents.

Related guides

AI changes the lure. These controls stop the outcome.

Email fraud, sign-in, AI tools and practice.

Get help your way

Six ways to reach us. Pick whichever suits you.

Helpdesk hours are Monday to Friday, 8am to 5pm AEST. On Advantage and above, emergencies get a response within 1 hour and outages within 2, with 24/7 best-effort escalation.

Questions

What people ask about AI threats.

How are deepfakes used against small businesses?

Usually as a short video or phone call that looks or sounds like a director, asking for an urgent payment, a bank detail change or a password reset. It is often followed by an email from a lookalike address. The defence is a rule that these requests are always checked on a separate channel.

Can scam emails still be spotted by bad spelling?

No. AI writes clean, convincing messages that can mention real projects and people found online. Judge requests by what they ask for, not how well they are written.

How much audio does someone need to clone a voice?

Only a short sample, such as a webinar, podcast or social video. Treat any voice instruction to move money, change bank details or reset access as unverified until checked.

What is prompt injection?

Hidden instructions in a document, email or web page that hijack an AI assistant when it reads them, for example to leak earlier conversation or draft a phishing reply. Only feed AI tools content you trust, and use business AI tools with logging.

Is there official guidance on AI agents?

Yes. On 1 May 2026, CISA, ASD's ACSC and Five Eyes partners published joint guidance on securing agentic AI: least privilege, human approval for high-impact actions, monitoring and logging, testing, and starting with low-risk uses.

What is the single most useful control?

A rule, written down and practised, that payments, bank detail changes, password resets and MFA changes are always verified on a channel you start yourself, no matter who seems to be asking.

Next step

Want your leadership team ready for a deepfake call?

We run AI-themed incident walkthroughs with leaders, set up the verification rules, and harden email and sign-in so one convincing call does not cost you money.

Talk to us 07 3114 2808