Australia Warns AI-Fueled Threats May Force Online Services Offline

The report said the agency responded to 1,500 cybersecurity incidents in the 2025-26 year, 20% more than the previous year, and notified businesses, governments and critical infrastructure operators of more than 4,400 potential malicious cyber incidents, an 80% increase. It said government and private networks experienced persistent breach attempts from state-based actors last year, and that actors in the Indo-Pacific could be launching attacks to prepare for future conflicts, including intelligence collection or disruption.

The ASD said hostile foreign governments, criminals and hackers are using AI to gather intelligence, disrupt critical infrastructure and steal money. Bradshaw said: “We need AI tools to help automate our defences to stay ahead of AI-enabled attackers.” She said Australia must accelerate modernisation of its technology and cybersecurity foundations. The agency said integrating AI into cybersecurity could help find weaknesses or suspicious behavior and improve responses.

Bradshaw has previously discussed the ASD’s access to Anthropic’s Mythos model, which the frontier lab said it could not release publicly because the tool could identify unknown flaws in IT systems, The Guardian reported. Australian government agencies believe access to such frontier models and collaboration with leading AI labs is critical. Defence minister Richard Marles said Australians must enhance cyber defences while harnessing AI’s benefits, and home affairs and cybersecurity minister Tony Burke said the government was positioning Australia to respond to evolving threats and opportunities from emerging technologies.

Bradshaw identified legacy technology—outdated systems or devices no longer supported with new updates—as among the biggest risks and most attractive targets for hackers. The ASD urged organisations to update systems, seek its help and report suspicious incidents. Bradshaw and Marles said organisations should prioritise security over keeping systems online, which could mean more services are forced offline while they are secured. The warning followed newly discovered vulnerabilities last week in PHP, a programming language that underpins most websites, which triggered a rush by governments and businesses to fix the problem. Bradshaw said: “In a threat environment transformed by AI, collective action is our strongest defence.”

A TechRadar column by a Saviynt strategy executive argued that chief information security officers and chief information officers should not chase hype about AI escaping sandboxes and morphing threats at the expense of identity security. The column said breaches still exploit missing essential controls around credentials and permissions, and that “AI washing”—overstating AI’s importance in a product, project or strategy—can obscure first principles. It cited the Hugging Face breach earlier this year, saying autonomous AI agents exploited an over-permissioned, unmonitored non-human identity that had cluster-admin access across two environments when compromised.

The column said human identities and non-human identities share the same failure mode: access is granted, never reviewed and never revoked. It said the identity challenge predates AI by decades, and that unclear ownership of AI identities and offboarding can leave gaps as organisations hire quickly and rely on freelancers, consultants, suppliers, auditors and temporary staff. It cited Verizon’s 2026 Data Breach Investigations Report, which found exploitation of vulnerabilities had overtaken stolen credentials as the top initial access vector for the first time in 19 years, while credentials still drove privilege escalation and lateral movement. The column argued that least-privilege access, just-in-time access and governance of non-human identities would have contained the Hugging Face breach after the vulnerability was exploited, and that identity problems cannot be patched away.