Phishing is entering a new phase marked by the use of generative artificial intelligence and advanced evasion and personalization techniques. Traditional phishing is a fraud based on easily identifiable emails or links, but phishing driven by AI causes the development of dynamic attacks capable of adapting in real time to the user, the context and the device. A recent study by Unit 42, the Palo Alto Networks threat intelligence team, warns about a new type of web attack in which seemingly benign pages are transformed, in a matter of seconds, into fully functional and personalized phishing sites.
In recent years, phishing has established itself as a persistent threat in Spain. Without going any further, last yeara wave of scams in La Rioja caused losses of more than 72,500 euros among several victims. Also, in 2025, these types of attacks were repeated in Catalonia, where an organization was dismantled that, through massive smishing campaigns, managed to steal 1.9 million euros to hundreds of people affected.
Similar cases occurred that same year in municipalities of Castilla y León and the Community of Madrid with losses that were more fragmented, but that reflect the effectiveness of phishing, demonstrating how it has ceased to be an isolated fraud and has become an everyday threat, distributed territorially and with an increasingly greater economic impact.
AI-powered phishing: more evasive, personalized and difficult to detect
Unlike traditional attacks, these AI-powered phishing pages do not contain detectable malicious code at the time of loading. Instead, they use client-side API calls to legitimate modeling services. large scale language (LLM) for generate malicious JavaScript snippets in real time and bypassing AI security barriers. After this process, the fragments are returned through the API, and are assembled and executed directly in the victim’s browser, leaving no trace of static loads that can be previously analyzed by conventional security solutions.
The result is a highly evasive attack assembled and executed during chargein which the code of the phishing page is polymorphic, with a single variant of the malicious code. Additionally, content is delivered from widely used and trusted LLM domains, allowing network analysis systems to be bypassed and reinforcing the effectiveness of deception.
Unit 42 states that 36% of pages exhibit run-time assembly behavior and that leveraging runtime LLMs on a web page allows attackers to bypass network analysis, increase the diversity of malicious scripts with each visit, use runtime assembly and execute JavaScript code to complicate detection, and obfuscate plaintext code.
A change in the fraud detection paradigm
The ability to generate fraudulent content in real time, personalize it, and execute it directly in the browser introduces a significant defense challenge. When malicious code does not exist until the moment of execution, traditional security approaches are insufficient.
Phishing driven by AI causes the development of dynamic attacks capable of adapting in real time to the user, the context and the device
Faced with this new scenario, Palo Alto Networks recommends:
- Bet on real-time behavior analysis during execution, capable of detecting suspicious activities at the moment they occur, regardless of their origin or legitimate appearance.
- Limit the use of unauthorized language modeling services in the work environment.
- Strengthen security barriers in the AI platforms themselves to prevent malicious use.
