AI Assistants Are Recommending Malicious Links When Users Ask for Legitimate Websites

A request for the T3 Code download leads to an apparent impersonator rather than the genuine product, according to an account posted by Dorian D. on X. He says his wife receives the link from ChatGPT and describes the destination as an almost pixel-perfect clone with a domain registered two weeks earlier.

Theo, who identifies himself as CEO of the businesses associated with T3 Code, takes up the report. After investigating throughout the day, he says ChatGPT repeatedly recommends the clone. An ordinary attempt to locate software has become a problem for the people behind the product: the assistant is directing users towards a site they identify as a scam.

Ordinary Website Requests Return Malicious Destinations

Testing by Netcraft, a security company that sells threat detection and takedown services, finds malicious recommendations beyond this reported encounter. Its June 2026 research tests ChatGPT, Copilot, Gemini and Perplexity with requests for the login pages and online services of major banking and retail brands.

Malicious links appear in 1.7% of the 2,905 responses generated. The destinations are live attacker-controlled infrastructure, rather than merely invented addresses or inactive websites.

The prompts resemble routine attempts to find a service, including questions about accessing a brand's account online. In one example, the company identifies a phishing page impersonating a bank's login portal as the first link in a Copilot response.

That placement matters because the user is already asking for the legitimate destination. A malicious page does not need to arrive through an unsolicited message if an assistant selects it while answering a request for the real service.

The Problem Extends Beyond Invented Addresses

An incorrect URL is not always an existing scam site. Netcraft's earlier research, published in July 2025, finds suggested domains that do not belong to the intended brands, including unregistered, parked or inactive addresses.

Such destinations create opportunities for someone to register or repurpose them. The 2026 findings describe a different exposure: malicious infrastructure that already exists and makes its way into generated answers.

Web-connected assistants can retrieve current material to inform their responses, rather than produce every answer solely from information learned during training. That gives them access to recent information, but also requires them to select sources from a web containing impersonators and phishing pages.

Netcraft identifies generative engine optimisation, or GEO, as a route for manipulation. Like attempts to influence conventional search rankings, the approach shapes content to gain visibility, but its target is inclusion in an AI-generated answer.

An assistant can therefore recommend a real, functioning website and still identify the wrong destination. The address does not have to be invented for the recommendation to be dangerous.

The Reported Recommendations Continue After the Site Goes Down

The route behind the T3 Code recommendations remains unclear. Theo's post makes that uncertainty explicit: 'I have no idea what these crypto scammers did, but their fake T3 Code scam clone gets recommended by ChatGPT constantly.'

His account describes repeated encounters, not a measured recommendation rate. The wider research establishes that malicious destinations appear in generated answers, but does not identify what makes ChatGPT select this particular clone.

Replies propose explanations involving search indexes and ranking manipulation. None supplies a verified account of how the impersonator enters the answers, and the supplied material contains no technical analysis of a downloaded file.

Theo subsequently reports a change in the destination's availability, but not in the assistant's recommendation. Taking a page offline removes access to that page; it does not, in his reported experience, immediately remove its address from ChatGPT's answers.

His update is brief: 'Site's down but ChatGPT is still recommending it 🙃'