Attackers Using LLMs to Create Phishing Pages in Real Time

“When the first webpage loads in the victim's browser, it requests client-side JavaScript from well-known and reliable LLM clients (such as DeepSeek and Google Gemini); however, the Proof of Concept may work with several models.”

 “Attackers can then use properly crafted prompts that get around security measures to fool the LLM into returning malicious JavaScript snippets. A completely functional phishing page is then produced by assembling and running these snippets in the browser's runtime. This eliminates any observable, static payload.”

 “Careful, quick engineering to get over the LLM's built-in defenses was essential to the attack's effectiveness.”

 “We discovered that straightforward rewording worked quite well. For example, whilst a direct request for "code to exfiltrate credentials" was prohibited, a request for a generic $AJAX POST function was allowed. To maintain the first page's cleanliness, indications of compromise (IoCs), such as Base64-encoded exfiltration URLs, could potentially be concealed inside the prompt itself.”

“This exploit is extremely difficult to defend against due to its dynamic nature and browser runtime assembly. For each victim, this attack mechanism generates a different variation. Every malicious payload is sent over a trusted domain and is dynamically produced and unique.”

Note: To get a stress-free working environment, you can go for a specially designed tool, “PhishNext,” that provides specialized simulations of phishing attacks so that the users can get used to such attacks and never become a victim of such attacks.