Study Maps LLM Role in Phishing Attacks

A systematic review of research from 2023 to 2025 examines how large language models are reshaping both phishing attacks and their detection.

Updated on Nov. 20, 2026 in Cybersecurity

Bold flat-color editorial illustration of a single fishing hook over a grid of identical cream envelope tiles, conveying the lure of AI-generated phishing at scale.
A systematic review of 2023-2025 research finds large language models let criminals produce convincing phishing emails and fake sites at scale, while AI also improves detection. AI Illustration. Upload story photo >

Live Poll

Do you feel more at risk of online scams because of AI tools?

Researchers from Cardiff Metropolitan University and international partners published a systematic literature review in the journal Array on large language models in phishing. The review analyzed research published between January 2023 and April 2025.

Why it matters

The review finds that large language models let criminals generate convincing, contextually relevant emails and fraudulent websites at scale while lowering the technical expertise needed to run campaigns. AI-powered detection systems can analyze content to identify phishing attempts more effectively than many traditional approaches.

The review covers research published between January 2023 and April 2025, a window that captures the rapid rise of large language models in cybercrime. The review does not quantify how much detection accuracy improves over traditional methods, and the extent of that gain remains to be measured across broader datasets.

The players

Cardiff Metropolitan University

A public university in Cardiff, Wales, whose Cybersecurity and Information Networks Centre led the review's authorship.

Dr Chaminda Hewage

A cybersecurity researcher at Cardiff Metropolitan University who co-authored the review.

Dr Rajkumar Rathore

A cybersecurity researcher at Cardiff Metropolitan University who co-authored the review.

The details

The review, co-authored by researchers from Cardiff Metropolitan University, the University of Greenwich, Pukyong National University, the University of Essex and the Beijing University of Posts and Telecommunications, examines how large language models automate phishing content creation. Cardiff Met's Cybersecurity and Information Networks Centre, led by Dr Chaminda Hewage and Dr Rajkumar Rathore, contributed to the work.

Timeline

  1. January 2023: Start of the research window analyzed in the review.

  2. April 2025: End of the research window analyzed in the review.

  3. 2026: Review published in the journal Array.

The Tech Race

The review follows the post-2022 surge in generative AI adoption, when large language models moved from research labs into widely available tools. It documents an arms race in which the same models are being used by attackers to scale phishing and by defenders to detect it.

Because large language models make convincing phishing emails easier to produce, readers should expect more personalized and believable scam messages in their inboxes. Practical defenses include verifying unexpected requests through a second channel and enabling multi-factor authentication on important accounts.

The takeaway

Large language models lower the barrier to running phishing campaigns, so treating unexpected messages with skepticism remains the most effective everyday defense. Organizations should pair that habit with AI-based detection tools that analyze content rather than relying only on sender reputation.

Further reading

For more on how criminals and defenders are adapting, explore our Cybersecurity coverage.

Source note: This article includes information reported by Business News Wales.

Live Poll

Do you feel more at risk of online scams because of AI tools?