CYBERSECURITY NEWS

AI-Powered Cyberattacks in 2026: What's Actually Happened, and How to Protect Yourself

Published 25 October 2026

Illustration of a circuit-board brain on a dark blue background

"AI hacks" are in the news constantly, and much of the coverage mixes real incidents with guesswork. This article sticks to what has been documented by the AI companies themselves, government agencies and established security researchers between mid-2025 and October 2026. Each item has a date and a source.

The short version: AI hasn't invented new kinds of attack. It has made the old ones (phishing, impersonation, malware, break-ins) cheaper, faster and more convincing. The defences that work are mostly the same ones that worked before, applied more consistently.

What's changed

Three shifts show up again and again in the official reports.

The skill bar has dropped. Anthropic's September 2026 threat report puts it bluntly: "Sophisticated attacks no longer require sophisticated attackers." It says AI has collapsed the gap between well-funded state hacking teams and individual operators.
Speed and scale have increased. Australia's ASD said in its Annual Cyber Threat Report 2024–25 (14 October 2025) that AI "almost certainly enables malicious cyber actors to execute attacks on a larger scale and at a faster rate" (iTWire coverage).
Scams are more convincing. When releasing 2025 scam figures on 30 March 2026, ACCC Deputy Chair Catriona Lowe said scammers "can create fake voices" and "fake videos that appear very convincing and are completely fabricated" (ABC News).

Notable incidents and trends, mid-2025 to October 2026

1. An AI ran most of a hacking campaign (November 2025)

On 13 November 2025, Anthropic reported that a group it assessed as Chinese state-sponsored had used its Claude Code tool to attack roughly 30 organisations, including tech companies, banks, chemical manufacturers and government agencies, in a campaign detected in mid-September 2025. Anthropic estimated the AI did 80–90% of the work, with humans stepping in at only a handful of decision points. The attackers "succeeded in a small number of cases".

It wasn't flawless: Anthropic noted the AI sometimes "hallucinated" credentials or claimed to have found secrets that were actually public. That's a useful reminder that these tools still make mistakes, for attackers as much as anyone.

2. Malware that asks an AI what to do next (June–November 2025)

Google's Threat Intelligence Group reported on 5 November 2025 that, for the first time, it had found malware families that call an AI model while they run. One, which Google tracks as PROMPTSTEAL, was used by the Russian government-backed group APT28 against Ukraine from June 2025; it asks a language model to generate commands, then runs them. Another, PROMPTFLUX, asks Google's Gemini to rewrite its own code to dodge antivirus. Google said PROMPTFLUX appeared to be still in testing and could not yet compromise a device, and that it disabled the assets involved.

3. AI agents doing "nearly all of the work" (December 2025 to August 2026)

Anthropic's September 2026 report, covering December 2025 to August 2026, describes a step up from the 2025 cases. A Russia-linked espionage actor whose tradecraft matched the group Microsoft tracks as Midnight Blizzard used AI agents to automatically rebuild its malware whenever security products flagged it, targeting more than 20 government, defence and diplomatic organisations (SecurityWeek).

Several cases touch ordinary phone users directly. One actor linked to the ShinyHunters crime group ran a pipeline that downloaded 1.8 million Android apps and scanned them for passwords and keys that developers had left inside (BleepingComputer, 11 September 2026). The report also describes actors hijacking WhatsApp accounts by quietly linking them as "companion devices", and compromising at least three companies that run hotel guest Wi-Fi to push Android and iOS malware to guests. Anthropic says it banned the accounts involved and shared what it found with authorities and industry.

4. Voice clones of real people (July 2025) and the fraud figures

In July 2025, a State Department cable reported that someone had used AI-generated voice and text messages, sent over Signal, to impersonate US Secretary of State Marco Rubio and contact at least three foreign ministers, a US governor and a member of Congress (CBS News, ABC News, 9 July 2025). The cable said the likely aim was to gain access to information or accounts.

The same technique is used on families. The FBI's 2025 Internet Crime Report, published in April 2026, logged 22,364 US complaints involving AI, with losses of more than US$893 million. It specifically mentions "distress" scams that use voice cloning to mimic a loved one in trouble, with over US$5 million lost to them in 2025. The FBI notes that investment scam losses overall exceeded US$8 billion, suggesting many victims never realised AI was involved.

5. AI browsers tricked by hidden instructions (August–December 2025)

"Agentic" browsers can read pages and take actions for you, which creates a new weakness called prompt injection: instructions hidden in a web page that the AI follows as if you'd typed them. On 20 August 2025, Brave published research showing that hidden text in a Reddit comment could make Perplexity's Comet browser fetch the user's email address and a one-time login code from their Gmail and send them to an attacker, all after the user simply asked it to summarise the page. Brave later said Perplexity's fix was incomplete.

On 22 December 2025, OpenAI wrote that prompt injection, "much like scams and social engineering on the web, is unlikely to ever be fully 'solved'", and advised users to limit what logged-in accounts agents can reach and to give them narrow, specific tasks (TechCrunch).

6. AI coding tools turned against their owners (August 2025)

On 26 August 2025, malicious versions of Nx, a popular developer package, were published. Besides stealing tokens directly, the malware tried to get AI command-line tools already installed on victims' computers (Claude, Gemini and Amazon Q) to hunt for sensitive files. Wiz found (3 September 2025) that the AI step succeeded in under a quarter of cases, partly because models refused, but the wider attack still exposed thousands of secrets and over 6,700 private code repositories.

7. AI is finding security holes faster, for both sides (2026)

On 7 April 2026, Anthropic announced Project Glasswing, saying an unreleased model had found thousands of high-severity vulnerabilities, including in every major operating system and web browser. It chose not to release the model publicly and is working with Apple, Google, Microsoft and others to fix the bugs. Its September report also describes a China-linked group using AI to find several previously unknown flaws in a major security product. The practical point for you: bugs are being found faster, so updates matter more than ever.

The Australian picture

The ACCC's National Anti-Scam Centre reported on 30 March 2026 that Australians lost $2.18 billion to scams in 2025, up 7.8% on 2024, with investment scams the largest category at $837.7 million (ACCC). For the first half of 2026, Scamwatch received 91,767 reports with $156.6 million lost (Information Age, 25 August 2026). Neither figure separates out AI-enabled scams, and ASD's 2025–26 threat report had not been published at the time of writing.

What it means for your phone

Your phone is where most of these attacks land, because it holds your messages, your banking, your authenticator codes and your contacts.

Phishing links: AI makes texts and emails fluent, personalised and free of the spelling mistakes people used to spot. Judge a message by what it asks you to do, not by how polished it looks.
Fake apps and leaky apps: Anthropic's 1.8-million-app scan shows attackers mining apps for developers' mistakes at scale. Install only what you need, from official stores, and delete apps you no longer use.
Voice-clone calls: A few seconds of audio from social media can be enough to imitate someone. A familiar voice on the phone is no longer proof of who's calling.
Zero-click exploits found faster: AI-assisted bug hunting means serious flaws in iOS and Android are being found, and potentially exploited, more quickly. Spyware such as Pegasus shows how damaging a single unpatched flaw can be.

If you're worried something is already on your phone, read how to check your phone for spyware or run our free in-browser spyware scanner.

Practical protection checklist

Verify by calling back. If a call, voice note or video asks for money, codes or urgent action, hang up and call the person or business on a number you already have. This defeats almost every voice-clone scam.
Agree on a family code word. Pick a word only your family knows and ask for it in any "emergency" call.
Use passkeys and multi-factor authentication. Passkeys can't be phished the way passwords can. Where you must use MFA, prefer an authenticator app over SMS codes, which are exposed to SIM swap scams.
Install updates promptly. Turn on automatic updates for your phone's operating system, browser and apps. With bugs being found faster, delays matter.
Turn on your phone's strongest protection mode. On iPhone, that's Lockdown Mode. On recent Android phones, it's Advanced Protection. Both reduce the ways an attacker can reach your device.
Don't paste secrets into AI tools. Passwords, recovery codes, ID documents and client data don't belong in chatbots or AI assistants.
Be cautious with AI browser agents. Don't let an AI agent loose while you're logged in to email, banking or work accounts. Give it narrow tasks and check before it sends, buys or shares anything.
Watch your linked devices. In WhatsApp and Signal, check Linked devices regularly and remove anything you don't recognise.
Report scams. In Australia, report to Scamwatch and, for cybercrime, to ReportCyber.

For a fuller walkthrough of settings, see our phone privacy checklist.

A calmer way to think about it

None of these incidents require panic. Most of them were caught and stopped, often by the AI companies themselves, and the defences above work against both the AI-assisted and the old-fashioned versions of each attack. The biggest risks remain the familiar ones: being rushed into acting on a message or call, and running outdated software.

Some people prefer to reduce their exposure further with a phone built around security. GrapheneOS, a hardened version of Android for Google Pixel phones, adds stronger app sandboxing, exploit protections and fine-grained permissions, and works without Google services. If that interests you, our PF Secured GrapheneOS phone comes set up and ready to use.

Note: this article summarises public reports available on 24 October 2026. Figures are as published by each source; we'll update it as new reports, including ASD's next Annual Cyber Threat Report, are released.

See the GrapheneOS phone→

← Back to Blogs