AI phishing emails: what the research shows and how to train
What the FBI says about generative AI in fraud, what a 2024 human-subjects study found about AI-written spear phishing, and how SecurityRing’s AI Builder drafts simulations you still approve.
Attackers now use generative AI to write phishing that looks like ordinary email. Defenders can use the same class of tools to practise those attacks safely. This post sticks to two primary sources: an FBI public service announcement on generative AI in fraud, and a 2024 human-subjects study on AI-automated spear phishing. Then it describes how SecurityRing’s AI Builder actually works.
Key takeaway
The FBI warns that generative AI makes fraud more believable and cheaper to scale. A 2024 study found AI-written spear-phishing emails matching human experts on click-through rate. Training has to practise personalised, current lures, not last year’s template.
What the FBI has warned about
On 3 December 2024 the FBI’s Internet Crime Complaint Center issued PSA I-120324-PSA, “Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud”. Among the examples it lists for AI-generated text:
- Messages sent faster to a wider audience with believable content.
- Language translation that reduces the grammar and spelling mistakes that once gave foreign actors away.
- Content for fraudulent websites and chatbots that prompt victims to click malicious links.
- Text used in social engineering and spear phishing.
The same PSA covers AI-generated images, voice cloning and videos used to impersonate executives, law enforcement and relatives. Those modalities matter for deepfake and vishing programmes; this post stays on email text.
What one research study measured
Heiding, Lermen, Kao, Schneier and Vishwanath (arXiv:2412.00586, 30 November 2024) sent four kinds of email to 101 recruited participants:
| Group | Click-through rate |
|---|---|
| Control (ordinary phishing / spam-style) | 12% |
| Human-expert spear phishing | 54% |
| Fully AI-automated spear phishing | 54% |
| AI with human-in-the-loop edits | 56% |
The AI tool scraped public information to build a profile, then generated and sent the email. The authors report that the AI-gathered information was accurate and useful in 88% of cases. This is one study with a recruited sample; it is not an industry benchmark and not a SecurityRing statistic.
What auditors look for
If you report AI-phishing readiness to the board, cite the primary source (date, sample size, click-through definition) and separate it from your own campaign metrics. Do not present a single study as “the” click rate for AI phishing.
How SecurityRing’s AI Builder works
SecurityRing uses AI on the defender side, with a human in control:
- Recommendations from your industry, tools, company context and current advisories, plus anonymised patterns from other SecurityRing customers’ recent campaigns in your peer industry.
- AI Builder drafts the brief, email, landing page, quiz and training video from a one-line brief or from “Build this idea” on a recommendation.
- You approve every tab before finalise. The red flags you mark become the lesson’s teaching points.
- Personalisation fills each employee’s first name, manager, department and location at send time.
- After a click, the employee sees a short video of that exact email with its red flags highlighted, and a quiz, designed to take under two minutes.
How SecurityRing helps
See AI phishing simulation for the product page. SecurityRing does not claim that AI-written simulations reduce breaches by a percentage. Across 100+ SecurityRing campaigns that have trained 10,000+ people, more than half of the people who clicked completed the in-the-moment lesson without a reminder.
Sources
Every clause, incident and figure above comes from these official or original sources.
- FBI IC3Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (I-120324-PSA)Published 3 December 2024 · Accessed 16 October 2026.
- arXiv (Heiding, Lermen, Kao, Schneier, Vishwanath)Evaluating Large Language Models’ Capability to Launch Fully Automated Spear Phishing Campaigns: Validated on Human SubjectsPublished 30 November 2024 · Accessed 16 October 2026.
