How AI is changing phishing, scams and website attacks

AI has not given criminals a new kind of attack. It has made the old ones faster, cheaper and more convincing. Phishing emails now arrive in fluent, well-targeted language, voices and video calls can be faked, and newly published software flaws are attacked within days. Government agencies and the AI companies themselves report the same picture. This article explains what has changed, what has not, and the defences that still work.

What AI actually changes

The UK National Cyber Security Centre (NCSC) assessed in 2024 that AI would almost certainly increase the volume and impact of cyber attacks, with the biggest gains in reconnaissance and social engineering. Its 2025 update found AI mainly enhancing existing methods rather than creating new ones.

The companies that make AI models agree. Google's threat intelligence team found criminals using its Gemini model for research, coding help and writing content: productivity gains, not new capabilities. OpenAI found scam networks fitting AI into "existing scam playbooks". Anthropic, which makes Claude, reported in 2026 that most attacks still use familiar methods such as stolen passwords, phishing and known flaws. What changed was speed and scale.

The other change is who can attack. The NCSC, Google and Anthropic all report that AI lowers the skill needed. In one case Anthropic described, a criminal with little technical skill used AI to build ransomware and sell it to others.

Phishing that reads perfectly

For years the advice was to look for spelling mistakes and awkward grammar. The NCSC warned in 2024 that generative AI produces lures without those errors. The FBI says criminals use AI to translate scam messages into flawless local language. Europol reports scams tailored to the victim's cultural context and personal details.

It works. Microsoft's 2025 Digital Defense Report found AI-enhanced phishing achieved a 54% click-through rate in its data, 4.5 times that of traditional phishing.

So check the things AI cannot fix:

• The real sender address, not the display name. Our help article on reading full email headers shows how.
• Where links actually go. Hover before you click.
• The request itself. Urgency, secrecy, a change of bank details or a request for a login code are the signs that matter, however well written the message is.

Our help article on identifying spam and phishing emails lists more.

Voices and faces you cannot trust on their own

The FBI warned in December 2024 that criminals use AI to clone the voices of relatives and public figures, and to hold real-time video calls posing as company executives. In 2025 it reported AI-generated voice messages impersonating senior officials to take over people's accounts, and noted that AI voices can sound nearly identical to the real person.

The defence is a habit, not a tool:

• Call back on a number you already have before acting on any request for money, a payment change or account access, however familiar the voice.
• Agree a code word with family or colleagues for urgent requests.
• Never give anyone a two-factor login code. No genuine bank, host or colleague needs it.

Websites under faster attack

The NCSC says the time between a software flaw being disclosed and attackers exploiting it has shrunk to days, and that AI will likely shrink it further. For the most attacked WordPress plugin flaws, the security company Patchstack measured it in hours. The NCSC also predicts a growing divide between systems that keep up with updates and those that do not.

Google has also reported the first malware that calls an AI model while it runs, to rewrite itself or choose commands. It describes this as early and experimental. The practical lesson is older: update your website software promptly, and remove what you do not use. Our article on why anyone would attack a small business website explains the methods.

New lures: fake AI tools

AI's popularity is itself a lure. Anthropic reported fake download sites spreading malware disguised as AI tools, and resellers offering cheap "AI access" that harvested people's logins. Criminals also steal AI service keys so they can run AI at the victim's expense.

Download AI apps only from the vendor's own website or a recognised app store, and treat an AI service key like a password.

AI on the defenders' side

The same technology helps defence. OpenAI estimates that ChatGPT is used to identify scams up to three times as often as it is used to run them. Google, OpenAI and Anthropic each publish reports on the misuse they find, ban the accounts involved and share what they learn with others. Spam filters have long learned from the mail people mark as spam, and they keep learning as campaigns change.

What we do on our side

At Noiz, we see the same worldwide campaigns as every mail and web host, and we answer them in layers:

• Email filtering on every hosted mailbox, tuned as new phishing campaigns appear, with risky attachment types blocked. The Email Security add-on adds advanced spam and virus filtering, spam training from what your users mark, and keeps SPF, DKIM and DMARC configured so criminals find it harder to send mail in your name.
• Faster patching for websites. WordPress Care updates WordPress, themes and plugins every week with a restore point first, and patches known plugin and theme vulnerabilities as fixes are released.
• A web application firewall and rate limits on our shared hosting that refuse known attack patterns and limit automated login attempts. Our AI, Bot and Crawler Policy explains how we treat AI crawlers: allowed, but within fair limits.
• Guidance on using AI safely. Our help article on using ChatGPT, Claude and Copilot safely with your email explains what to watch.

Five habits that still work

1. Check the request, not the writing. Polished language proves nothing.
2. Call back on a number you already have before moving money or changing payment details.
3. Use two-factor authentication everywhere, and never share a code.
4. Keep website software and computers updated, promptly.
5. Protect your own domain with SPF, DKIM and DMARC, so criminals cannot easily use it to fool your clients.

If you would like your domain's email checked and protected, add Email Security from the console, or open a ticket and we will look at it with you.

Sources

Checked 19 October 2026.

Government and agency assessments

• NCSC, The near-term impact of AI on the cyber threat (January 2024)
• NCSC, Impact of AI on cyber threat from now to 2027 (May 2025)
• ENISA Threat Landscape 2025 (press release)
• FBI IC3, Criminals use generative AI to facilitate financial fraud (December 2024)
• FBI IC3, Senior US officials impersonated in malicious messaging campaign (May 2025)
• European Commission, summary of Europol's IOCTA 2025

Platform and AI company reports

• Microsoft, Digital Defense Report 2025 release
• Google Threat Intelligence Group, Adversarial misuse of generative AI (January 2025)
• Google Threat Intelligence Group, AI threat tracker (November 2025)
• OpenAI, Disrupting malicious uses of AI: an update (October 2025, PDF)
• Anthropic, Detecting and countering misuse of AI (August 2025)
• Anthropic, Threat intelligence report (September 2026)
• Patchstack, State of WordPress Security in 2026

Noiz

• Email Security add-on
• WordPress Care
• AI, Bot and Crawler Policy
• Help articles linked in the text, on docs.noiz.ie

Note: while some of this content was produced with the help of AI, the final editing has been done by a real human being.