ChatGPT search links New Zealanders to unsafe sites

ESET has warned that ChatGPT search results are directing New Zealand users to unsafe websites, including scam sites, fake login pages and malicious downloads.

The warning comes as AI tools are used more widely in New Zealand to find products, software and information online. ESET says four out of five New Zealanders now use AI and describes ChatGPT as one of the country's most widely used consumer apps.

ESET researchers monitored ChatGPT results over a 30-day period and found links to websites already classified as unsafe. Examples included fake online stores, cryptocurrency scams and imitation login pages. Some responses also linked to malicious computer code that could steal information or compromise a user's device.

Scott Leman, Country Manager, ESET NZ, said the findings challenge a growing assumption that links suggested by an AI assistant have already been checked for safety.

“Kiwis are increasingly using AI to tell them where to buy something, what software to download or which website to visit. The risk is that we start placing blind faith in the answers these tools provide. Someone asks an AI tool for help, follows the link it provides and ends up on a fake login page, scam website or malicious download. Because that recommendation has come from a tool they trust, they may be less likely to question whether the destination itself is genuine or safe. What this data shows is that users cannot assume a link is safe simply because ChatGPT has provided it,” Leman said.

Leman said AI-generated links should be treated with the same caution as links received through email, text messages or social media. A fake Microsoft login page or imitation retail site could lead users to hand over usernames, passwords or payment details, or to download harmful files.

Phishing rise

Separate ESET data showed phishing-related threats made up 35% of all cyber threats detected across its New Zealand user base in July. Phishing attacks are designed to trick people into clicking links, opening files, scanning QR codes or entering sensitive information into websites that appear legitimate.

“If the destination is a fake Microsoft login page or an imitation retail site, the consequences can be very real. A user could enter their username and password, provide financial information or download something harmful without realising they have been directed to a fraudulent site,” Leman said.

He said growing AI use is changing both how people find information online and how attackers exploit trust. That shift is reducing the usefulness of some traditional warning signs, such as poor spelling or awkward grammar.

“We have reached a tipping point. AI isn't simply making phishing emails more convincing, it is also changing the way people find and trust information online. An AI-generated answer is not automatically a security endorsement of the website, application or download it recommends,” Leman said.

Faster attacks

ESET also said AI is cutting the time between the public disclosure of a software vulnerability and its use by attackers. Leman said businesses once had time to understand a newly discovered flaw, install a patch and reduce their exposure.

That window is now shrinking sharply. Citing analysis from Zero Day Clock, a global project tracking software vulnerabilities confirmed as exploited in real-world attacks, ESET said the median period between public disclosure and exploitation had fallen from 771 days in 2018 to zero days in 2026.

“For years businesses could reasonably assume that when a new security weakness was discovered there would be some time to understand it, fix it and protect themselves. That breathing space is disappearing. AI can now analyse a newly released security patch, work backwards to identify the flaw it was designed to fix and help turn that information into a working attack extremely quickly. In one recent test involving a browser security update, an advanced AI model was able to analyse the patch and produce a working exploit in less than an hour. That is work which historically required specialist expertise and considerably more time,” Leman said.

A zero-day median means attackers are typically exploiting weaknesses by the time they become publicly known, leaving organisations little or no opportunity to install a security update before attacks begin. Leman said that is especially difficult for smaller businesses, which often lack dedicated security teams and round-the-clock monitoring.

Account access

ESET said another common misconception is that antivirus software on its own is enough to stop modern attacks. Leman said many phishing incidents do not rely on a virus file at all, but instead use a convincing fake page to capture valid login credentials.

Once an attacker has access to a genuine email account, the damage can spread quickly. They may be able to monitor conversations, access files, impersonate the account holder and use the compromised account to target colleagues, customers or suppliers.

“Antivirus remains an essential first line of defence, but not every modern cyber attack involves a virus. If someone receives what looks like a genuine Microsoft 365 message, follows a link and enters their username and password into a convincing fake login page, the attacker may then have legitimate credentials. There may be no virus file for antivirus software to find because the criminal is effectively walking through the front door using the user's own key,” Leman said.

He said that access could then be used to alter payment instructions within an existing email chain, making invoice fraud harder to spot because the message appears to come from a real account.

“An attacker could see that an invoice is about to be paid and then send another email from the genuine account saying the bank details have changed. The customer has little reason to suspect anything is wrong because the message has come from the same email address and may even form part of an existing conversation. By the time either side realises what has happened, the money may already be gone,” Leman said.

“When attackers were moving at human speed, delays were already a problem. As AI increasingly allows attacks to operate at machine speed, businesses and consumers can no longer assume they will have time on their side,” Leman said.