Phishing attacks continue to be a pervasive threat in the digital landscape. With almost a million unique phishing web pages appearing in the first quarter of 2024 alone, traditional defenses are struggling to keep up. Blacklists are quickly outdated, and even sophisticated machine learning models can be easily evaded by increasingly sophisticated phishing techniques that mimic legitimate sites. Researchers are now turning to powerful multimodal Large Language Models (LLMs) to tackle this challenge. LLMs, trained on massive amounts of data, can analyze not just the text of a webpage, but also its visual elements like logos, themes, and favicons to identify the brand a page is impersonating. In a new study, researchers built a two-phase system using LLMs. The first phase uses the LLM to identify the brand being represented, while the second phase verifies whether the website domain matches the identified brand. This two-pronged approach helps distinguish between, say, a real PayPal page and a fake one hosted on a suspicious domain. The research team tested GPT-4, Gemini Pro 1.0, and Claude3, and found that all three showed promising results, particularly when provided with both visual screenshots and HTML data. GPT-4 and Claude3 performed especially well, achieving high accuracy in detecting phishing attempts while also offering explanations for their decisions. Remarkably, these LLM-based systems outperformed existing state-of-the-art brand-based phishing detectors, showing greater resilience against adversarial attacks designed to fool AI. While the use of LLMs for security introduces new opportunities, it also presents new challenges. The cost of using these powerful models for real-time analysis, the potential for economic denial-of-service attacks, and the risk of indirect prompt injection are all important considerations. The accessibility of these LLMs also means attackers can study them to develop new evasion techniques. However, this research highlights the potential of multimodal LLMs as a significant step forward in the fight against online phishing.