Adaptive AI Sentinels Against Phishing Attacks: Democratizing Cybersecurity Through Interactive Learning

Abstract

Phishing attacks have become more convincing as generative AI enables attackers to create polished,
context-aware emails that closely resemble legitimate communication. These messages often evade traditional filters that rely on surface features and leave users without a clear understanding of why a message may be harmful. This work introduces an adaptive phishing-detection system that uses natural language processing to model semantic, linguistic, and stylistic signals and produce a risk score indicating how phish-like or benign an email appears. A complementary large language model layer then performs contextual and intent-based reasoning to interpret the deeper meaning of the message and detect subtle social engineering cues. The system incorporates adversarial and prompt-safety checks to strengthen reliability against AI-generated threats and through a web app, it delivers short micro-lessons for each detection, helping users understand the psychological tactics involved and learn to recognize them in future messages. This research contributes to both cybersecurity and NLP by showing how semantic scoring and LLM-based reasoning can be operationalized together to counter
AI-enabled social engineering while remaining interpretable for non-expert users. By combining accurate detection with continuous user education, the proposed solution strengthens trust, awareness, and long-term resilience, offering a scalable defense mechanism for modern phishing attacks.