# AI-Generated Phishing Emails Outperform Human Attackers by 60%, New Research Confirms Worst Fears


The cybersecurity community has watched the convergence of artificial intelligence and social engineering with mounting concern. Now, a comprehensive study from Proofpoint and Hoxhunt has quantified what security professionals suspected: AI-generated phishing campaigns are not merely competitive with human-crafted attacks—they demonstrably outperform them by a substantial margin, creating an asymmetric advantage that shows no signs of diminishing.


## The Threat


Employees clicked on AI-generated phishing emails at rates 60 percent higher than those crafted by experienced human threat actors, according to the joint research conducted throughout the final two quarters of 2024. In spear-phishing scenarios where attackers target specific individuals or departments, the performance gap widened further: AI-generated emails achieved a 73 percent improvement in open rates and subsequent click-through rates compared to their human-authored counterparts.


These findings emerge from the analysis of over 1.5 million phishing simulation emails deployed across 100 organizations—a statistically robust sample that captures behavior across industries, geographies, and employee demographics. The sheer scale of the research eliminates the possibility that these results represent an outlier or artifact of testing methodology. Instead, they represent a fundamental shift in the threat landscape: attackers who leverage AI for campaign composition now possess a measurable advantage over those who rely on traditional social engineering tradecraft.


## Background and Context


Phishing remains the entry point for the vast majority of enterprise breaches. Threat actors understand this well. Over the past three years, the barrier to entry for launching sophisticated phishing campaigns has steadily declined—not because attacks have become simpler, but because attackers now have access to automation tools that handle the most labor-intensive aspects of social engineering.


The emergence of large language models and agentic AI systems created new possibilities for threat actors. Unlike traditional phishing campaigns, which required human attackers to manually research targets, craft messaging, and distribute emails at manageable scale, modern AI systems can automate nearly every step. The question was never whether threat actors *would* adopt these tools, but how quickly and to what effect.


The Proofpoint-Hoxhunt research provides the answer: adoption is happening, and the effect is pronounced.


## How the Research Was Conducted


The study's methodology reflects industry best practices in controlled phishing simulations. Researchers recruited experienced human social engineers—professionals with extensive track records in legitimate security awareness training—to craft one set of phishing emails. A second set was generated by AI systems, which included fine-tuned large language models and agentic AI pipelines capable of gathering real-time intelligence before composing each message.


The AI systems had access to diverse data streams. LinkedIn activity, corporate websites, employee directories, social media profiles, and public news sources all fed into the generation pipeline. For each target, the AI system researched recent professional activities, recent company announcements, and contextual information that would allow the phishing email to reference timely, credible details. This approach transformed phishing from a generic broadcast attack into a highly contextualized, individually tailored social engineering attempt.


Both sets of emails—human-crafted and AI-generated—were deployed as simulations to real employees across the 100 participating organizations. The metrics captured were straightforward: Did the employee click the link? Did they submit credentials or other sensitive information?


## Why AI Phishing Succeeds


Several distinct advantages explain AI's superiority in this arms race:


Hyper-personalization at scale represents the most obvious advantage. Human social engineers face practical constraints: they can realistically personalize a handful of spear-phishing emails per day. AI systems generate thousands of personalized emails simultaneously, each calibrated for a specific target's role, recent activities, and organizational context.


Real-time context injection allows AI phishing to reference events that occurred hours or even minutes before the email was sent. An AI system monitoring a company's recent hiring announcements can reference a new department head by name and title. It can mention a recent product launch that appeared on the company's social media. These details create immediate credibility and urgency, triggering the cognitive shortcuts that humans rely upon when processing business communications under time pressure.


Linguistic precision in matching expected business communication patterns presents another advantage. LLMs trained on millions of business emails understand the conventions of corporate communication—the level of formality, the expected structure, the vocabulary appropriate to each department and industry. An AI-generated email requesting credential updates from the finance department reads indistinguishably from legitimate internal communications, whereas human-written phishing often contains subtle grammatical tells or tonal inconsistencies that trained employees can identify.


Language independence eliminates a traditional barrier to global phishing campaigns. AI systems generate native, culturally appropriate phishing emails in any language with equal proficiency. Threat actors can now scale campaigns across multiple regions without the coordination overhead or quality inconsistency that characterizes human-driven, multilingual operations.


## Real-World Threat Deployment


The research is not merely theoretical. Proofpoint has identified multiple criminal organizations that have operationalized these capabilities. Several threat groups now operate Phishing-as-a-Service (PhaaS) platforms that integrate AI generation pipelines into their service offerings.


One documented platform charges $150 monthly for access to an AI system capable of generating and distributing 10,000 personalized phishing emails per day. This pricing—within reach of individual cybercriminals and small organized crime groups—democratizes access to sophisticated phishing infrastructure that would have required substantial technical expertise and investment just two years ago.


## Implications for Enterprise Defense


This research demands a reassessment of how organizations approach phishing risk. Traditional metrics—particularly click rates on phishing simulations—no longer provide reliable guidance for measuring employee readiness. If AI-generated phishing is significantly more persuasive than the simulations many organizations currently deploy, then employees trained on human-crafted phishing may possess false confidence in their ability to identify actual threats.


The findings also underscore that technical controls must become primary defense mechanisms. Behavioral change and employee awareness, while necessary, are insufficient against a threat class where the human cognitive processing advantage has fundamentally shifted toward the attacker.


## Recommendations


Organizations should prioritize a multi-layered approach that shifts weight toward technical controls:


Authentication infrastructure should move beyond passwords and SMS-based verification. Hardware security keys implementing FIDO2 standards and passkey protocols offer protection that remains effective even when credentials are compromised through phishing.


Privileged access management (PAM) solutions that enforce step-up authentication for sensitive operations—particularly financial transactions and system credential requests—introduce procedural friction that disrupts even perfectly crafted phishing attempts.


Employee training should evolve beyond grammar-spotting and signature analysis. Instead, focus on behavioral cues—unexpected requests for immediate action, pressure to bypass normal approval workflows, or communications that deviate from established processes. Teach call-back verification: when an email requests credentials or approves financial transfers, employees should independently verify the request using known contact information.


## HackWire Analysis


This research marks a turning point in the phishing arms race. For two decades, the field operated under the assumption that sufficiently trained employees could identify social engineering attempts through linguistic or behavioral analysis. That assumption no longer holds when attackers can generate phishing emails that match or exceed the linguistic sophistication of legitimate business communications.


The real concern is not that AI phishing is now effective—it is that we are witnessing the beginning of a capability expansion curve. Current AI systems represent the baseline. Future iterations will incorporate additional context layers, behavioral modeling, and even real-time response analysis. Organizations that continue to rely primarily on employee awareness and traditional technical controls are building their defenses against yesterday's threat, not tomorrow's.


Human-centric security measures remain valuable but can no longer serve as primary defense. The asymmetry that now favors attackers demands a fundamental rebalancing toward zero-trust authentication, technical enforcement, and architectural defenses that operate independently of employee decision-making under pressure.