AI brain analyzing a phishing email with human analysis magnifying glass.

AI vs. Human Analysts: Unmasking the Truth Behind Large Language Models' Accuracy in Cybersecurity

"Can AI Truly Replace Human Expertise in Analyzing Phishing Attacks? A Deep Dive into LLMs' Capabilities and Limitations"


Large Language Models (LLMs) have revolutionized various fields with their impressive ability to generate human-quality text and code. While LLMs excel at tasks like composing emails and essays, their capacity for statistically-driven descriptive analysis, particularly on user-specific data, remains largely unexplored. This is especially true for users with limited background knowledge seeking domain-specific insights.

This article delves into the accuracy of LLMs, specifically Generative Pre-trained Transformers (GPTs), in performing descriptive analysis within the cybersecurity domain. We examine their resilience and limitations in identifying hidden patterns and relationships within a dataset of phishing emails.

We explore whether these reasoning engine tools (LLMs) can be used as generative AI-based personal assistants to aid users with minimal or limited background knowledge in an application domain to carry out basic, as well as advanced statistical and domain-specific analysis. By comparing LLM-generated results with analyses performed by human cybersecurity experts, we aim to provide a clear understanding of AI's current capabilities and the continued importance of human expertise.

AI Search Multiple angles on this topic

Defining the Scale of the Challenge

The term 'large' is commonly defined as exceeding most other things of like kind in quantity or size, a descriptor increasingly relevant to modern language models whose parameter counts and training datasets dwarf those of earlier systems. In general usage, 'large' can describe anything above average in size or number, from consumer goods to computational architectures. As language models grow in scale, questions arise about whether sheer size translates into reliable analytical performance in high-stakes domains such as cybersecurity.

Established Methodologies and Their Constraints

Traditional cybersecurity analysis relies on established frameworks such as threat modeling, signature-based detection, and expert-driven incident response, each with documented strengths and limitations. These methods, while time-tested, often require significant manual effort and are constrained by the availability of trained personnel. As threat landscapes evolve, the scalability and speed of conventional approaches have come under scrutiny.

The Evolution of AI in Security Analysis

Early applications of artificial intelligence to cybersecurity focused on rule-based systems and statistical anomaly detection, which laid the groundwork for more sophisticated approaches. The emergence of neural networks and later deep learning marked a significant shift, enabling models to identify patterns without explicit programming. These milestones have progressively shaped expectations for what automated systems might achieve in threat identification and response.

LLMs vs. Human Analysts: A Comparative Analysis of Phishing Email Detection

AI brain analyzing a phishing email with human analysis magnifying glass.

This study investigates the effectiveness and precision of LLMs in data transformation, visualization, and statistical analysis on user-specific data. Unlike models trained on general datasets, this research focuses on LLMs' ability to analyze data not included in their original training set. It involves descriptive statistical analysis and Natural Language Processing (NLP)-based investigations on a dataset of phishing emails.

The experimental setup involved analyzing phishing emails using both human analysts and LLMs (specifically LangChain and GPT-4). Human analysts utilized mainstream tools and libraries like Python and NLTK. LLMs were tasked with the same analysis, allowing for a direct comparison of their performance.

Key areas of comparison include:|Feature Engineering: LLMs and analysts were compared on their ability to extract key features from email subject lines and bodies, such as length, word count, and verb/noun counts.|Emotional Affect Analysis: The study assessed the ability of LLMs and human analysts to identify and interpret the emotional tone of phishing emails, utilizing tools like NRCLex.|Correlation Matrix Generation: Both LLMs and human analysts were tasked with creating correlation matrices to identify relationships between different variables within the phishing email data.|Temporal Analysis: The research examined how well LLMs and human analysts could identify patterns in phishing email activity over time (e.g., peak months, days, and hours for attacks).
AI Search Multiple angles on this topic

Contemporary Studies on LLM Performance

Recent research has begun evaluating large language models on cybersecurity-specific tasks, with studies examining capabilities in vulnerability detection, threat classification, and incident summarization. Results so far suggest that these models can sometimes match or approach human-level accuracy on narrowly defined benchmarks, though performance varies significantly by task type. The field remains in an early phase, with limited large-scale, peer-reviewed comparative studies available.

Documented Shortcomings and Failure Modes

Critics have highlighted several failure modes of large language models in security contexts, including susceptibility to prompt injection, generation of plausible but incorrect threat assessments, and inability to reason about novel attack vectors. Studies have noted that these models can produce confident-sounding outputs that lack factual grounding, posing risks in operational environments where accuracy is critical. The absence of robust uncertainty quantification further complicates their deployment alongside human analysts.

Human vs. Model: Measuring Relative Performance

Comparative studies between human analysts and AI systems often focus on metrics such as detection accuracy, false positive rates, and time-to-identification. In certain standardized tasks, large language models have demonstrated competitive performance, particularly when processing large volumes of unstructured text data. However, human analysts retain clear advantages in contextual reasoning, ethical judgment, and adaptability to unprecedented threat scenarios.

While LLMs demonstrated strong performance in numerical reasoning tasks and feature engineering, they encountered difficulties with domain-specific analysis. For instance, GPT-4 struggled to classify the polarity (sentiment) of phishing emails and generate correlation matrices from textual data. This is because GPT-4 was not able to use specific libraries. Human analysts are still needed in some cases.

The Future of AI in Cybersecurity Analysis

LLMs are transforming cybersecurity. Although this has been studied, in the future these new tools must be researched further for all areas of human jobs. We can improve emotional analysis and correlations with added libraries to create domain-specific algorithms.

AI Search Multiple angles on this topic

Integrating Findings and Expert Perspectives

Synthesizing available evidence suggests that large language models offer promising augmentative capabilities for cybersecurity analysts rather than wholesale replacements. Experts generally caution against over-reliance on automated outputs without human oversight, particularly in high-consequence decision-making contexts. The most effective approaches appear to involve collaborative workflows that leverage the speed and scale of AI alongside human expertise and judgment.

Emerging Directions and Open Questions

Future research is expected to focus on improving the reliability, interpretability, and domain-specific fine-tuning of large language models for security applications. Areas of active investigation include better calibration of model confidence, integration with real-time threat intelligence feeds, and development of robust evaluation benchmarks. As models continue to evolve, the balance of responsibilities between human and machine analysts will likely shift in ways that remain difficult to predict.

Wider Implications and Structural Barriers

The adoption of AI in cybersecurity exists within broader systemic challenges, including workforce shortages, regulatory uncertainty, and the escalating complexity of global threat landscapes. Organizations face difficulties in integrating new AI tools with legacy infrastructure and existing security operations center workflows. These structural barriers may moderate the pace at which large language models are deployed at scale, regardless of their technical capabilities.

Practical Consequences for Analysts and Organizations

In practice, the introduction of AI tools into cybersecurity teams has both augmented and disrupted traditional analyst roles, with effects varying by organizational context. Real-world deployments have demonstrated gains in efficiency for routine tasks such as log analysis and report generation, while more complex investigative work continues to depend on human judgment. The ultimate impact on cybersecurity outcomes will likely depend on how effectively organizations manage the transition to hybrid human-AI operational models.

About this Article -

Written with AI assistance from published research, and reviewed by the Mystum team. See our About page for more information.

This article is based on research published under:

DOI-LINK: https://doi.org/10.48550/arXiv.2405.19578,

Title: The Accuracy Of Domain Specific And Descriptive Analysis Generated By Large Language Models

Subject: cs.ce econ.gn q-fin.ec

Authors: Denish Omondi Otieno, Faranak Abri, Sima Siami-Namini, Akbar Siami Namin

Published: 29-05-2024

Everything You Need To Know

1

What are the core differences in how LLMs and human analysts approach phishing email detection?

LLMs, particularly GPT-4, are tested for their capabilities in several areas. These include feature engineering (extracting characteristics like length and word count), emotional affect analysis (identifying sentiment using tools like NRCLex), correlation matrix generation (finding relationships between variables), and temporal analysis (identifying patterns over time). Human analysts, on the other hand, employ established tools and libraries such as Python and NLTK. The study compares the effectiveness and precision of LLMs against human experts in these critical areas of cybersecurity analysis.

2

What specific challenges did GPT-4 face when analyzing phishing emails, and why?

GPT-4 encountered difficulties in several domain-specific tasks. It struggled to accurately classify the polarity (sentiment) of phishing emails and to generate correlation matrices from textual data. This limitation stems from GPT-4's inability to directly utilize specific libraries, hindering its capacity to perform in-depth, specialized analysis that human analysts can achieve with tools like Python and NLTK. This highlights the current dependence on human expertise for complex, nuanced cybersecurity investigations.

3

In what areas do LLMs like GPT-4 excel, and what does this imply for the future of cybersecurity?

LLMs demonstrated strong performance in numerical reasoning tasks and feature engineering, showcasing their potential in streamlining certain aspects of cybersecurity analysis. This suggests that LLMs can be effectively used as AI-based personal assistants to aid users with minimal background knowledge to carry out basic as well as advanced statistical and domain-specific analysis. However, the limitations in domain-specific analysis highlight the need for continued human oversight and the potential for hybrid approaches that combine AI strengths with human expertise.

4

How was the research conducted to compare the performance of LLMs and human analysts in phishing email detection?

The research employed a direct comparison. The experimental setup involved having both human analysts and LLMs (specifically GPT-4 and LangChain) analyze the same dataset of phishing emails. Human analysts used established tools and libraries like Python and NLTK. The LLMs were tasked with the same analyses. The comparison focused on feature engineering, emotional affect analysis, correlation matrix generation, and temporal analysis to assess their respective capabilities and limitations in this critical cybersecurity task.

5

What are the key takeaways from this comparison of AI and human analysts in phishing email detection, and what future developments are anticipated?

The study reveals that while LLMs show promise in certain areas, human expertise remains essential for comprehensive cybersecurity analysis. GPT-4's difficulties with domain-specific tasks like sentiment analysis and correlation matrix generation emphasize the need for specialized tools and human understanding. Future developments include incorporating specialized libraries to enhance emotional analysis and create domain-specific algorithms, potentially enabling LLMs to better support human analysts. The study stresses the importance of combining the strengths of both LLMs and human analysts to enhance cybersecurity effectiveness.

Newsletter Subscribe

Subscribe to get the latest articles and insights directly in your inbox.