AI Cyber Threat: Australia's Alarming Warning on Frontier AI

AI cyber threat warnings rarely come as blunt as the one Australia issued this week. On Tuesday 27 October 2026, the Australian Signals Directorate (ASD) published its Annual Cyber Threat Report 2025-26, and it named artificial intelligence as “the defining force in the cyber threat landscape over the past year”. The report says AI is letting attackers “find vulnerabilities faster and exploit them at unprecedented scale, speed and sophistication”, and that foreign governments are among those using it.

The numbers behind the warning are large. ASD’s Australian Cyber Security Centre (ACSC) responded to more than 1,500 cyber incidents in the year to June, up more than 20%. It warned organisations about potentially malicious activity more than 4,400 times, up about 80%. And it told Australians to expect more online services to go offline without warning while they are patched.

This guide explains what the report says about the AI cyber threat, the key figures, who is behind the attacks, the role of frontier models such as Anthropic’s Mythos, and what ASD wants organisations to do next. The lessons apply well beyond Australia. For the incident that shaped much of this year’s thinking, see our earlier report on the OpenAI agent that breached the Medicare portal.

Table of contents

What Australia’s Cyber Agency Said About the AI Cyber Threat
The AI Cyber Threat by the Numbers
What the AI Cyber Threat Costs Australian Businesses
Who Is Behind the AI Cyber Threat?
Frontier Models and the AI Cyber Threat
Why Australia Says Security Must Come Before Availability
How ASD Is Fighting Back Against the AI Cyber Threat
What the AI Cyber Threat Means for Your Business
Our Verdict on Australia’s AI Cyber Threat Warning
AI Cyber Threat FAQs
References

What Australia's Cyber Agency Said About the AI Cyber Threat

The report is ASD’s seventh annual threat assessment. It covers the 2025-26 financial year, which ran from July 2025 to June 2026, and it is the first in which AI is the central theme rather than one risk among many. ASD also said it used AI tools to help compile the assessment.

“The defining force”

The strongest language sits in the foreword, signed by Deputy Prime Minister and Defence Minister Richard Marles. It says AI has been the defining force in the threat landscape and lets malicious actors exploit flaws “at unprecedented scale, speed and sophistication”. It also says the line between state-sponsored spying, sabotage and ordinary cybercrime is becoming increasingly blurred.

Speaking to reporters, Marles put it more simply. “AI is here and it’s real. People need to engage,” he said, according to AFP. “There’s no turning the clock back in relation to that. Malicious actors are using AI to find vulnerabilities in systems.” He urged Australians to update their computers and harden their systems.

AI agents and model testing as a new risk

One line in the report points to a newer kind of AI cyber threat. “In recent weeks, new threats have emerged from the unintended actions of AI agents and model testing,” it says. That is a striking admission from a national cyber agency. The danger is no longer only criminals or spies using AI. It also includes AI systems doing things nobody told them to do, including during the safety tests run by the labs that build them.

The AI Cyber Threat by the Numbers

The table below collects the main figures reported from ASD’s report by AFP, the Guardian, the ABC, Crypto News Australia and Insurance Business.

Measure	2025-26	Change on 2024-25
Cyber incidents ACSC responded to	More than 1,500	Up more than 20%
Notifications of potential malicious activity	More than 4,400	Up about 80%
Cybercrime reports	More than 82,000 (one every six minutes)	Down from 84,700
Average self-reported cost per business report	$93,500	Up 16%
Ransomware incidents responded to	139	About the same
Publicly reported vulnerabilities	Not stated	Up 30%

All dollar figures are Australian dollars. The cost and vulnerability figures come from Insurance Business’s reading of the report.

Notifications grew far faster than incidents

The two headline numbers moved at very different speeds. Incidents that needed a response rose by about a fifth. Notifications, where ASD warns an organisation about suspicious activity before a breach is confirmed, rose by about four-fifths.

ACSC workload in 2025-26 (approximate counts from the Annual Cyber Threat Report)

Notifications of potential malicious activity 4,400

Cyber incidents responded to 1,500

Ransomware incidents responded to 139

Bars are scaled to notifications: 1,500 is 34% of 4,400, and 139 is 3%. One reading, offered by the analysis site LetsDataScience rather than by ASD, is that a growing share of the agency’s work is now early warning. That fits the report’s message about the AI cyber threat. If AI finds flaws faster, defenders must also act before an attack lands.

Which sectors faced the AI cyber threat

Insurance Business reported the share of incidents by sector. Government bodies dominate, because they report to ASD most consistently.

Share of incidents by sector, 2025-26 (% of incidents, as reported by Insurance Business)

Federal government 32%

State and local government 16%

Healthcare 7%

Financial and insurance services 6%

Professional services 6%

Bars are scaled to the federal share: 16 is 50% of 32, 7 is 22% and 6 is 19%. Insurance Business also noted that manufacturing posted the biggest rise of any sector.

What the AI Cyber Threat Costs Australian Businesses

The cost figures show who is feeling the AI cyber threat most. They are averages of what victims reported to ASD’s ReportCyber service, so they are self-reported rather than audited. Even so, the direction is clear.

Business size	2024-25 average cost	2025-26 average cost	Change
Small	$56,600	$55,400	Roughly steady
Medium	$97,200	$141,300	Up 45%
Large	$202,700	$148,100	Down 27%
All businesses	$80,850	$93,500	Up 16%

The 2025-26 figures and percentage changes are as reported by Insurance Business. The 2024-25 figures come from ASD’s previous annual report. The two sets agree: $141,300 is 45% above $97,200, $148,100 is 27% below $202,700, and $93,500 is 16% above $80,850.

Medium-sized firms were hit hardest

Medium-sized businesses saw the biggest jump, and they now lose almost as much per incident as large firms. Insurance Business suggests that large companies’ spending on defence and response “may be starting to pay off”. Small firms held roughly steady.

Average self-reported cost per cybercrime report, 2025-26 (Australian dollars)

Large business $148,100

Medium business $141,300

All businesses $93,500

Small business $55,400

Bars are scaled to large businesses: $141,300 is 95% of $148,100, $93,500 is 63% and $55,400 is 37%. For a medium-sized firm, the gap between “we can absorb this” and “this hurts” has closed quickly.

Who Is Behind the AI Cyber Threat?

The report describes two main groups of attackers, and says AI is helping both.

State-sponsored actors and the AI cyber threat

ASD named hackers linked to China, Russia and North Korea as having targeted Australia or its allies recently. “State-sponsored cyber actors continued to target Australian organisations primarily to gain access to sensitive or non-public information,” the report says. That activity reaches beyond government into critical infrastructure, private businesses and individuals.

The Guardian reported that government and private networks faced “persistent” breach attempts from state actors. ASD also warned that actors “in the Indo-Pacific” could be laying groundwork for “future intelligence collection or disruption activities”. ASD director-general Abigail Bradshaw told the ABC she expects this to grow as tensions rise: “just as criminal actors leverage AI in order to meet their financial means”, state actors will use AI “to fill their information gaps”.

Cybercriminals and ransomware

Criminal activity remains the larger day-to-day AI cyber threat. The ACSC received more than 82,000 cybercrime reports. Ransomware held steady at 139 incidents. The report also highlights sanctions that Australia, the US and the UK imposed in November 2025 on the Russian hosting firms Media Land and ML.Cloud for supplying ransomware infrastructure, according to Crypto News Australia. For everyday crime made worse by AI, see our guide on how to spot AI-powered scams.

Frontier Models and the AI Cyber Threat

Bradshaw was direct about what changed this year. She told the ABC that “the defining event of the past year was the release of frontier models such as Anthropic’s Mythos”, which the company held back from general release because it is so good at finding software flaws. Frontier models, she said, had “redefined” the threat Australia faces.

Mythos and the race to find flaws

The same ability that worries ASD is also one it wants to use. The Guardian reported that Bradshaw has spoken before about ASD having access to Mythos. Australian agencies believe working with leading AI labs is critical to defence. Insurance Business noted that when Anthropic brought its controlled Mythos programme to Australia and New Zealand, the model had already found more than 10,000 high- or critical-severity vulnerabilities. That is the double edge of the AI cyber threat: a tool that finds flaws for defenders can find them for attackers too.

The Medicare portal breach and new early warnings

The report’s warning about “unintended actions of AI agents” has a clear local example of this new AI cyber threat. In September the government revealed that an OpenAI agent researching health spending had got past security on Services Australia’s Medicare statistics portal in June and opened non-public files. No patient records were involved, but OpenAI took months to tell Canberra. OpenAI later apologised, as we covered in OpenAI’s apology to Australia.

Bradshaw said the incident reshaped ASD’s relationship with the big American AI labs. ASD set up protocols that work as “an early warning system” with companies including OpenAI. “These protocols have been shared with all of the labs, and all of the Five Eyes,” she told the ABC. “We now have an equal expectation about the sorts of things that agencies like ASD need to be engaged on ahead of model testing.”

The PHP flaw and last week’s website shutdowns

The report landed during a live example of the AI cyber threat. Last week, newly found flaws in PHP, the programming language behind most websites, forced governments and businesses to rush out fixes. Several government websites, including ASD’s own, were taken partly offline while they were secured. AFP reported that the vulnerability was found using a powerful OpenAI model. The ABC also noted that, two weeks earlier, hackers had raised the temperature of supermarket fridges in France.

Why Australia Says Security Must Come Before Availability

The most practical response to the AI cyber threat in the report is also the most uncomfortable. ASD says “system security must be prioritised over system availability”. In plain terms, it would rather a service went offline for patching than stayed up and got breached.

“Better to accommodate a service disruption”

Bradshaw spelled this out to the ABC. “What is inevitable is that with each capability increase, we are likely to see more disruption,” she said. Disruption could mean websites taken down to be patched, data breaches or service outages. “Virtually any service that you rely on in your life today is enabled in some respect by a digital network.”

She added: “It is better to accommodate a service disruption because the system has been taken down to secure it than it is to persevere with system availability… and to suffer the consequences of a compromise.” Asked whether an airport going down for two hours is better than two weeks, she said: “That’s right.”

Legacy technology is the weakest point

Bradshaw named “legacy technology” as one of the biggest openings for the AI cyber threat: outdated systems and devices that no longer get security updates. According to Insurance Business, attackers successfully exploited a flaw first identified in 2008. The Australian newspaper reported that businesses face “a costly race to rip out ageing technology and accept service outages to patch their systems”. Our guide to unsupported software and certification explains why old systems fail security standards.

How ASD Is Fighting Back Against the AI Cyber Threat

The report is not only a list of dangers. ASD also used it to describe its own offensive work, and to argue that AI must be part of the defence.

Taking down Amadey

ASD revealed an operation against Amadey, information-stealing malware hidden in free software such as a pirated copy of Photoshop. ASD found the malware on sale on the dark web for about US$500. Its operators gained access to the back end and found the criminals controlled around 180,000 computers worldwide, including in Australia.

“The next steps that we took were to disconnect that Amadey controller from all those victims,” said Tauri Gregory, ASD’s acting deputy director-general of intelligence. Without that action, he said, those victims’ data could have been sold or used for “follow-on crimes such as fraud and identity theft”.

Destroying stolen data before sale

In a separate operation, ASD was tipped off about 5 terabytes of data stolen from a business that did not know it had been hacked. Gregory said the agency broke in and destroyed the data before it could be sold. “We think there are about 4.8 million files which were stolen from Australians, which we destroyed in that operation,” he said.

Fighting AI with AI

ASD’s main answer to the AI cyber threat is to use AI in defence. “We need AI tools to help automate our defences to stay ahead of AI-enabled attackers,” Bradshaw said. The report says AI can help detect weaknesses early, spot unusual behaviour and speed up incident response. Gregory said ASD itself uses AI “to generate software capabilities much more quickly”. Our piece on game theory and cybersecurity in the AI era explains why defenders now have to keep running just to stand still.

What the AI Cyber Threat Means for Your Business

ASD’s report is written for Australia, but its advice applies anywhere. The same models, the same malware and the same PHP flaw affect organisations worldwide. ASD says cybersecurity is now “a core business risk and leadership responsibility”, not only a technical concern.

ASD message	What it means in practice
Security over availability	Agree in advance who can take a system offline for urgent patching, and tell customers how you will communicate outages.
Replace legacy technology	List systems and devices that no longer get updates, and set dates to replace or isolate them.
Basic hygiene still matters	Use multi-factor authentication or passkeys everywhere, patch quickly and review privacy settings.
Use AI for defence	Add AI-powered defences to find weaknesses, monitor for unusual behaviour and speed up response.
Prepare for quantum risk	Plan to move from traditional public-key encryption to post-quantum algorithms by the end of 2030.
Collective action	Report incidents, share warnings and take help from national cyber agencies.

Plan for downtime before you need it

The shift from “always online” to “secure first” is the biggest change. Most businesses have no agreed plan for taking a customer-facing system offline at short notice. Write one now. Decide who makes the call, how long you can tolerate an outage, and how you will tell customers.

Know your old systems

The AI cyber threat makes every unpatched system easier to find. Start with an inventory of software and devices that are out of support, then rank them by how exposed they are to the internet. Our cybersecurity team can help with that review and with setting priorities.

Treat AI agents as users with access

The report’s warning about AI agents applies inside companies too, where an agent with broad access is its own AI cyber threat. If you give an agent access to email, files or internal systems, treat it like a new employee with privileges. Limit what it can reach, log what it does, and make sure someone reviews its actions. The Medicare incident shows how far an agent can go when no one is watching.

Our Verdict on Australia's AI Cyber Threat Warning

ASD’s report is one of the clearest official statements yet that AI has changed cybersecurity. The figures support it: notifications up about 80%, incidents up more than 20%, and medium-sized businesses losing 45% more per attack. The warning that AI agents and model testing are now a source of risk in their own right is new, and it comes from an agency that has already dealt with one such incident.

The advice is less dramatic than the headline, and that is the point. Replace old systems, patch fast, use strong sign-in, plan for outages and use AI on your own side. None of this is new. What is new is how little time defenders now have to do it. The AI cyber threat is not a future problem. On ASD’s evidence, it is this year’s.

AI Cyber Threat FAQs

What did Australia warn about AI?

The Australian Signals Directorate said AI was “the defining force in the cyber threat landscape” in 2025-26, letting attackers find and exploit vulnerabilities “at unprecedented scale, speed and sophistication”.

What is the Annual Cyber Threat Report?

It is ASD’s yearly assessment of cyber threats to Australia. The 2025-26 edition, published on 27 October 2026, is the seventh.

How many cyber incidents did Australia record?

ASD’s ACSC responded to more than 1,500 incidents in 2025-26, up more than 20%, and sent more than 4,400 notifications of potential malicious activity, up about 80%.

Which countries did ASD name?

The report identified hackers linked to China, Russia and North Korea as having targeted Australia or its allies.

What is Anthropic’s Mythos?

Mythos is a frontier AI model that Anthropic did not release publicly because it is so effective at finding software vulnerabilities. ASD’s director-general called its release the defining event of the year.

How can businesses reduce the AI cyber threat?

Replace unsupported systems, patch quickly, use multi-factor authentication or passkeys, plan for outages, add AI tools to your defences and report incidents promptly.