Large Language Models in Cybersecurity: State-of-the-Art
Farzad Nourmohammadzadeh Motlagh
1
, Mehrdad Hajizadeh
2
, Mehryar Majd
1
, Pejman Najafi
1
,
Feng Cheng
1
and Christoph Meinel
1
1
Hasso-Plattner-Institute for Digital Engineering, University of Potsdam, Germany
2
Technische Universitat Chemnitz, Germany
Keywords:
LLM, Large Language Model, AI, Cybersecurity, Advanced Threats, Cyberattacks, Cyberdefense, Privacy
and Security.
Abstract:
The rise of Large Language Models (LLMs) has revolutionized our comprehension of intelligence bringing
us closer to Artificial Intelligence. Since their introduction, researchers have actively explored the applica-
tions of LLMs across diverse fields, significantly elevating capabilities. Cybersecurity, traditionally resistant
to data-driven solutions and slow to embrace machine learning, stands out as a domain. This study examines
the existing literature, providing a thorough characterization of both defensive and adversarial applications of
LLMs within the realm of cybersecurity. Our review not only surveys and categorizes the current landscape
but also identifies critical research gaps. By evaluating both offensive and defensive applications, we aim to
provide a holistic understanding of the potential risks and opportunities associated with LLM-driven cyberse-
curity.
1 INTRODUCTION
The evolution of generative artificial intelligence, no-
tably large language models (LLMs), has influenced
most disciplines of science and technology that sup-
port content generation in diverse applications (Neu-
pane et al., 2023). In education, LLMs support ed-
ucators in various tasks such as assignment assess-
ment (Hsiao et al., 2023), question generation (Elkins
et al., 2023), providing feedback (Guo and Wang,
2023), and essay grading (Yan et al., 2023). In the
entertainment industry, LLMs demonstrate competi-
tive performance in generating music captions (Deng
et al., 2023b) as well as video game scripts (La-
touche et al., 2023). Automation is introduced into
customer service (Pandya and Holia, 2023), market-
ing (Gan et al., 2023; Yang et al., 2023b), and sup-
ply chain management (Hendriksen, 2023; Li et al.,
2023a; Kosasih et al., 2023) through the integration
of LLMs in business. Meanwhile, the utilization of
LLMs in healthcare enables professionals by provid-
ing real-time clinical decision support (Rao et al.,
2023; Fawzi, 2023), medical education (Kuckelman
et al., 2023; Song et al., 2023), and prediction of
disease progression (Shoham and Rappoport, 2023;
Rasmy et al., 2021).
With advancements in cyber threats, the cyberse-
curity domain can also be equipped with cutting-edge
tools, assisting cybersecurity practitioners who con-
tinuously seek solutions to implement advanced poli-
cies or strengthen technological protections against
the disclosure of confidential information, unautho-
rized access, and other forms of data modification
(Kaur et al., 2023). Thanks to LLMs’ capability in
breaking down complex natural language patterns, se-
curity experts are now enabled to explore more at-
tack vectors in various contexts associated with tex-
tual data (Yang et al., 2023a).
Functionalities of LLMs are increasingly being
integrated into the cybersecurity posture, contribut-
ing to promising enhancements in cybersecurity de-
fense applications (Li et al., 2023b). Through ana-
lyzing vast amounts of text data, including security
logs, these models can identify emerging vulnerabili-
ties. Anomaly detection represents a key application
of LLMs for identifying potential threats (Liu et al.,
2023). Furthermore, LLMs mitigate potential risks
by offering automated vulnerability fixes, aiming to
improve organizations’ security posture (Pearce et al.,
2023).
However, with the continuous advancements of
LLMs in cyber defense, it is crucial to acknowledge
98
Motlagh, F. N., Hajizadeh, M., Majd, M., Najafi, P., Cheng, F. and Meinel, C.
Large Language Models in Cybersecurity: State-of-the-Art.
DOI: 10.5220/0013377600003899
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 11th International Conference on Information Systems Security and Privacy (ICISSP 2025) - Volume 2, pages 98-110
ISBN: 978-989-758-735-1; ISSN: 2184-4356
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

that these language models can also be leveraged by
malicious actors. For example, LLMs can be mis-
used by attackers to execute malware in target com-
panies (Botacin, 2023), engage in defense evasion
(Chatzoglou et al., 2023), and gain access to creden-
tials (Rando et al., 2023). The potential to generate
complex and personalized phishing messages further
highlights the misuse of LLMs for deceiving people
in an organization, paving the way for unauthorized
access to companies’ sensitive information (Saha Roy
et al., 2023; Jiang, 2024). To further elaborate, Wor-
mGPT (Falade, 2023) is an AI-powered tool designed
for cybercriminals to automate the generation of per-
sonalized phishing emails. Although it may sound
somewhat similar to ChatGPT, WormGPT is not a
friendly neighborhood AI; instead, its purpose is to
produce malicious content. Furthermore, FraudGPT
(Dutta, 2023) enabled attacker to create content to
convince users to click on a particular generated link.
The dual nature of LLMs has transformed the cy-
bersecurity realm by offering new challenges and op-
portunities. Developing robust defensive strategies to
foresee attacks and address concerns related to the uti-
lization of LLMs motivated us to formulate a taxon-
omy of strategies appearing in the field of cybersecu-
rity. To define our contributions more precisely, this
paper addresses:
• The intersection of LLMs’ offensive approaches
as a newly introduced dimension to cybersecurity
is framed in this study in line with the Mitre attack
framework (Corporation, 2023).
• Exploring LLM-empowered defensive strategies
in dealing with potential threats and malware
based on the NIST cybersecurity framework (Cy-
bersecurity, 2014).
• Understanding the major functionalities of LLMs
in current research trends alongside potential ap-
plications in the cybersecurity landscape.
The rest of paper is organized as follows: In Section
2, we provide an overview of LLMs . Moving forward
to Section 3, we explore cyber threat defenses levere-
gred by LLMs where Section 4 outlines sophisticated
attacks designed by LLMs. Finally, Section 5 con-
cludes the challenges posed by LLMs in the context
of cybersecurity.
2 BACKGROUND
LLMs are neural networks that learn from textual data
to process various language-related tasks (Naveed
et al., 2023). From Eliza as a pattern recognition chat-
bot in the 1960s (Weizenbaum, 1966), over the years
several advancements pushed Natural Language Pro-
cessing (NLP) forward, such as long short-term mem-
ories to handle a wide range of data (Hochreiter and
Schmidhuber, 1997), Stanford CoreNLP suite (Man-
ning et al., 2014) providing a collection of algorithms
to perform intricate NLP tasks and continued with
transformer architecture (Vaswani et al., 2017).
A breakthrough in Transformer-based models
surged the field of NLP and led to the development
of numerous kinds of effective LLMs. T5 (Raffel
et al., 2020) applied language modeling in pre-trained
LLMs, where spans are altered with a single mask.
GPT-3 enhanced the performance of LLMs with size
by increasing model parameters to 175B. PaLM-2 is
trained on high-quality datasets (Anil et al., 2023)
with an objective of cutting the cost of training and in-
ference (Naveed et al., 2023). Llama, a set of decoder-
only models aimed at minimizing the amount of ac-
tivations in the backward step (Naveed et al., 2023;
Touvron et al., 2023). Xuan Yuan 2.0, a Chinese fi-
nancial chat model (Naveed et al., 2023; Zhang and
Yang, 2023), AlexaTM (Soltan et al., 2022), PaLM-
2 (Anil et al., 2023), as well as GLM-130B (Zeng
et al., 2022) are a few instances of general purpose
pre-trained LLMs. While pre-trained models offer
an essential understanding of languages, as AI ad-
vances, fine-tuning LLMs boost business functions
and satisfaction by fulfilling industry-specific criteria
(Zhang et al., 2023b). A general-purpose LLaMA-
GPT-4 (Peng et al., 2023), Goat (Liu and Low,
2023) for handling complicated arithmetic queries,
HuaTuo (Wang et al., 2023) a medical knowledge
model, Evol-Instruct (Xu et al., 2023) offering com-
plicated prompts, and LLaMA 2-Chat fine-tuned us-
ing rejection sampling (Touvron et al., 2023) are
exemplary instruction-tuning LLMs. Running in
higher costs, extensive hardware requirements, cost
of slow training on various tasks, limited LLMs uti-
lization (Naveed et al., 2023). Retrieving support ev-
idence from an external in-domain knowledge base
(Zhang et al., 2023a), parameter tuning and knowl-
edge distillation are among the techniques extensively
researched for effective LLM deployment (Naveed
et al., 2023).
Recently, the scientific literature has experienced
a significant growth in the number of articles related
to LLMs, principally driven by their proven efficacy
across a wide range of functions. As a result, through-
out various surveys, researchers attempted to catego-
rize these advancements in LLM architecture (Naveed
et al., 2023; Zhao et al., 2023; Zhou et al., 2023;
Huang and Chang, 2022). Though previous studies
have investigated literature reviews to highlight the
safety aspects of LLMs (Iturbe et al., 2023; Adding-
Large Language Models in Cybersecurity: State-of-the-Art
99

ton, 2023; Kucharavy et al., 2023; Ishihara, 2023), the
present study focuses primarily on the application of
LLMs in the context of cyberdefense as well as cyber-
attack.
3 DEFENSIVE APPLICATIONS
OF LLMs
In the field of cybersecurity, the National Institute of
Standards and Technology (NIST) provides a com-
prehensive structure to enhance organizations’ cyber-
security status, as detailed in the NIST cybersecurity
framework (Cybersecurity, 2014). According to its
effectiveness and popularity in cyberdefense, we clas-
sify the diverse array of LLM-centered approaches
that contributed in cyberdefense through the lens of
NIST framework to better understand the impact of
LLMs in cyberdefense. The framework consists of
a structured approach to identify, protect, detect, re-
spond to, and recover from cybersecurity threats and
incidents.
3.1 Identify
The process of developing an organizational under-
standing to manage cybersecurity risk concerning
systems, assets, data, and capabilities is referred to the
Identify function in the context of the NIST frame-
work (Cybersecurity, 2014). Identifying potential
risks is a crucial phase in risk management, and LLMs
aim to fulfill a transformational role in forming risk
management in businesses. Johnson (Johnson, 2023)
presents invaluable insights for policymakers on the
applicability of LLMs to risk management. Accord-
ing to the author, LLMs go through business head-
lines, social media posts, economic indicators, le-
gal documentation, and other key sources, emphasiz-
ing risk elements to deliver more accurate and pre-
dictive risk assessments that a human analyst might
overlook. Lima et al. (de Lima et al., 2023) de-
velop a risk matrix from application reviews using
LLMs. Through user feedback, they proposed an au-
tomatic prompt extraction technique. These prompts
were passed into LLMs, which classified the risks
into five classes ranging from negligible to critical for
further investigation. Naleszkiewicz (Naleszkiewicz,
2023) discusses LLM applications allowing compa-
nies to overcome traditional enterprise risk manage-
ment challenges, such as operational and compliance
risks. LLMs evaluate unstructured siloed data across
various departments, acting as a bridge to provide an
in-depth understanding of an organization’s risk pro-
file. Furthermore, LLMs boost risk modeling by gen-
erating expert opinions based on prior patterns, risk
mitigation by generating contingency plans, and risk
reporting by providing customized risk assessments.
3.2 Protect
Implementing safeguards to guarantee the delivery of
essential services is reflected in protect function (Cy-
bersecurity, 2014). It involves various mechanisms
such as maintaining a proactive security posture or
prioritizing cybersecurity awareness and training to
empower the organization’s workforce. In the cur-
rent digital environment, proactive protection tech-
nologies are essential since they enable companies to
anticipate and prevent troubles before they arise. For
example, proactive technologies empower enterprises
to minimize the likelihood of coming across inappro-
priate content, and thus reduce the possibility of expe-
riencing ethical or legal challenges (Sun et al., 2023).
Voros et al. (V
¨
or
¨
os et al., 2023) harnessed the power
of LLMs to enhance web content filtration. They have
improved the accuracy of web content categorization
by scanning of large amount of URLs. Another re-
search accomplished by Yu et al. (Yu and Martin,
2023) investigates GPT-3’s capacity to produce hon-
eywords to trap the attackers if they are using decep-
tive generated passwords. First, they extract the com-
ponents of the original password using a password-
specific segmentation algorithm. These segments are
then fed into GPT-3 as a prompt to generate a collec-
tion of passwords similar to the input password. A
crucial element in this model’s efficacy is the mainte-
nance of strong password components called chunks
given to the LLM (Sannihith Lingutla, 2023).
LLMs can play a valuable role in strengthening
cybersecurity awareness and training within the pro-
tect function of the NIST framework. Tann et al.
(Tann et al., 2023) apply LLMs to tackle profes-
sional certification topics and perform Capture The
Flag (CTF) tasks to improve participants’ cybersecu-
rity education. LLMs have significance by enabling
attendees to explore CTF test settings, providing ex-
planations to concerns connected to professional cer-
tification, and highlighting the need to model cyberse-
curity breach scenarios in CTF sessions to support the
development of more comprehensive skills. However,
LLMs face limitations when it comes to responding to
conceptual queries. Furthermore, LLMs can improve
team collaboration by offering security question so-
lutions that are suitable for inexperienced as well as
experts. For instance, LLMs greatly increase the ef-
ficacy of penetration test teams by making it easier
for team members to pass on information by offer-
ing more in-depth assessments and generating appro-
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
100

priate explanations to be on the same page about the
detected risks. Moreover, LLMs serve as a connec-
tion between experts and publicly accessible web re-
sources, in particular assisting specialists in remain-
ing up to date on the most recent security concerns
that are critical to their company (Dutta et al., 2018).
Automated vulnerability fixing with LLMs dimin-
ishes the risk of cyberattacks. A three-step process
is described by Charalambous et al. (Charalambous
et al., 2023) for addressing automotive vulnerability
issues. Bounded Model Checking (BMC) is the first
step in the process. It evaluates the user-provided
source code to a property specification. The origi-
nal code and the appropriate counterexample are pro-
vided to the LLM module by the BMC engine in
the scenario that this phase’s verification is unsuc-
cessful and a security property violation is detected.
Secondly, customized queries are sent to the LLM
engine to produce a corrected version of the code.
Lastly, the BMC module re-evaluates the code that the
LLM module changed to formally determine whether
the updated version matches the original security and
safety requirements.
Automating flaw mitigation can be facilitated by
LLMs if the defect is well-defined and the prompt
provides additional information. While these mod-
els were fully effective in fixing simulated vulner-
abilities, real-world scenarios presented challenges
for their performance. The primary challenges stem
from the numerous methods that information is pre-
sented, the complexities of prompt processing and
code development in LLMs, and the significance of
prompt phrasing, which can result in notable vari-
ations in the code required to be generated (Pearce
et al., 2023). Furthermore, Sandoval et al. (San-
doval et al., 2023) performed an examination of po-
tentially insecure code suggestions during the process
of code development. Within a particular program-
ming context that the authors had defined, they tested
scenarios with and without AI support. Their findings
indicate that users assisted by AI develop security
flaws at a rate lower than ten percent, suggesting that
using LLMs in their security-oriented research does
not present major new security risks. Additionally,
Yu et al. (Fengrui and Du, 2024) present a method
for automating Tactics, Techniques, and Procedures
(TTP) classification in few-shot learning scenarios.
The method employs ChatGPT for data augmentation
and Instruction-Supervised Fine-Tuning on large lan-
guage models. Using ChatGPT results in diverse sam-
ple expansion that do not undermine the original text’s
contextual semantic.
3.3 Detect
The NIST framework’s Detect function serves to
identify cybersecurity events as they arise (Cyberse-
curity, 2014). Exploring anomaly detection in system
logs is a crucial step toward developing effective de-
tection methods through the use of LLMs. Recurrent
Neural Network Language Models are used by Tuor
et al. (Tuor et al., 2018) to present an unsupervised,
online anomaly detection method for computer secu-
rity log analysis. This approach simplifies the usual
effort-intensive feature engineering stage, making it
fast to implement, and is independent of the tools used
for system configuration and monitoring. The authors
have demonstrated the efficacy of their approach by
utilizing the Los Alamos National Laboratory Cyber
Security Dataset (Kent, 2016). Their findings indicate
that the approach can be handled in real-time, gen-
erating and organizing log-line-level anomaly scores
while taking into account inter-log-line context. The
authors (Tuor et al., 2018) considered metrics includ-
ing Average Percentile (AP) and Area under the Re-
ceiver Operator Characteristic Curve (AUC) to show
how the false-positive rate dropped without signifi-
cantly affecting the ability to detect unusual behavior
(Kent, 2016).
GPT-2 is used by VulDetect(Omar and Shiae-
les, 2023), a transformer-based vulnerability detec-
tion framework, to detect anomalies in system logs.
Using a dataset containing both vulnerable and non-
vulnerable code, the model is fine-tuned to detect
anomalies that represent regular behavior. Malicious
behavior is defined as any unexpected or unlikely out-
come that the model possibly generated. Two bench-
mark datasets, SARD (Zhou and Verma, 2022) and
SeVC (Shoeybi et al., 2019), were utilized by the au-
thors to assess VulDetect’s performance. The out-
comes showed that VulDetect has a low false positive
rate and is efficient in real-time vulnerability detec-
tion. Moreover, the integration of LLMs into pen-
etration testing practices has the potential to revolu-
tionize the world of threat detection. Threat detec-
tion could undergo a revolution if LLMs are incor-
porated into penetration testing procedures. Happe et
al.’s investigation (Happe and Cito, 2023) focused on
using LLMs to improve penetration testing. In line
with their classification, LLMs provide advancement
in two aspects of penetration testing: high-level and
low-level operations. High-level assignments include
conceptual investigation and strategic planning, such
as finding out about emerging active directory attacks.
On the other hand, tasks at a lower level incorporate
consideration of practical activities involving system
exploitation and vulnerability analysis. This entails
Large Language Models in Cybersecurity: State-of-the-Art
101

looking for specific attack vectors for a particular sys-
tem.
A further investigation by Deng et al. (Deng et al.,
2023a) introduces PENTESTGPT, an automated pen-
etration testing system driven by LLMs. Complex
tasks such as question answering, summarization, and
reasoning are readily handled with PENTESTGPT.
Addressing context loss concerns and simulating hu-
man behavior in penetration testing are the objec-
tives. Three self-interacting modules jointly form
PENTESTGPT including reasoning, generation, and
parsing. These modules collaborate to tackle penetra-
tion testing problems by using a divide-and-conquer
approach. Specific subtasks are allocated to each
module, which interact to effectively handle and com-
pile the data generated during testing.
Ranade et al. (Ranade et al., 2021) improve
the processing of threats, attacks, and vulnerabili-
ties which is challenging due to the high volume
of data, and the dynamic nature of evolving attack
techniques. The primary objective of their research
is an enhanced version of a BERT model, which
aims to effectively perform several cybersecurity-
related operations. Using Masked Language Model-
ing (MLM), the model was trained using unstructured
and semi-structured open-source Cyber Threat Intelli-
gence (CTI) data. Its evaluation encompassed diverse
downstream tasks with potential applications in Se-
curity Operations Centers (SOCs). They additionally
offer real-world examples of how to apply CyBERT
to cybersecurity problems. Several subsequent works
have furthered the advancements of this research in
terms of both training efficiency and accuracy such
as SecureBERT (Aghaei et al., 2022), CySecBERT
(Bayer et al., 2022), and ClaimsBERT (Ameri et al.,
2022). In this regard, Bayer et al. (Bayer et al., 2022)
presented a word embedding model based on BERT
and collected a dataset from multiple sources. This
adaptation makes the model capable of coping with
a wide range of cybersecurity tasks, namely malware
detection, alert aggregation, and phishing website de-
tection. Similarly, LILAC (Jiang et al., 2024) is a
log parsing method that employs an adaptive parsing
cache to boost the efficiency of log analysis proce-
dures. LILAC attempts to tackle issues such as in-
consistent outputs and a lack of specialized log pars-
ing capacities by updating templates using LLMs’ in-
context learning (ICL) power and a novel adaptive
parsing cache.
The LLMs can also facilitate auditory tasks to de-
tect vulnerabilities among the smart contracts. David
et al. (David et al., 2023) utilized LLMs to target vul-
nerabilities in the smart contracts and DeFi protocol
layers. Their study detects 52 compromised DeFi pro-
tocols, as input data for the language model context,
evaluating the impact of model temperature and con-
text length on the language model’s efficacy in smart
contract auditing. The results indicated that incor-
porating LLMs into the audit workflow substantially
boost the effectiveness and accuracy of analyzing an
array of feasible attacks. On the other hand, Chen et
al. (Chen et al., 2023) trained LLM on a dataset of
10,000 smart contracts and evaluated how well it de-
tected nine different vulnerabilities. According to the
authors’ findings, LLMs frequently deliver false posi-
tive results when detecting smart contract vulnerabil-
ities. This might be connected with interference from
incomplete codes or LLMs’ incapacity to understand
code segments.
An LLM can be used to build a scenario compa-
rable to an attacker’s strategy for gaining access to an
organization’s property by exploiting a vulnerability.
Garvey et al. (Garvey and Svendsen, 2023) study the
viability of using Generative-AI to improve the de-
velopment of Red Team scenarios in organizations.
The authors (Garvey and Svendsen, 2023) propose
employing LLMs to construct narratives based on
prompts or questions as input. Subsequently, subject-
matter specialists provide remarks, including modi-
fying narratives, adding new elements, or integrating
multiple items to develop more complex scenarios.
The objective is to guarantee that the generated sce-
narios are plausible and adhere to the provided frame-
work. They found that including elements inspired by
fiction into LLMs improves creativity and imagina-
tion in the scenario development process.
Koide et al. (Koide et al., 2023) present a strategy
for detecting phishing websites using LLMs. Their
approach entails using a web crawler to retrieve data
from websites and creating prompts for LLMs. Social
engineering strategies are then identified by evaluat-
ing the context of entire web pages and URLs. The
prompts rely on the Chain of Thought (CoT) prompt-
ing technique, which enables LLMs to elaborate on
their reasoning. In addition, the study recommends an
HTML simplification approach to improve efficiency.
This entails lowering the token count by simplifying
HTML text and removing HTML elements that lack
text within tags, such as style, script, and comment
tags. This operation is repeated until the token count
reaches a certain threshold, thus boosting overall effi-
ciency.
Sakaoglu introduced KARTAL(Sakaoglu, 2023),
a fine-tuned Language Model for detecting vulnera-
bilities in web applications. A detector component
in the KARTAL system is controlled by the prompts
from the prompter component. These prompts are
generated based on input gathered by the fuzzer com-
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
102

ponent, which monitors application activity. The
LLM detects logical vulnerabilities in web applica-
tions, specifically broken access control rules, by an-
alyzing these prompts. This technique allows KAR-
TAL to dynamically alter the definitions of broken
access, allowing it to adapt to a variety of scenarios.
This adaptability distinguishes it from less intelligent
vulnerability scanners, allowing KARTAL to be more
effective in its detection capabilities.
LLMs demonstrate their capacity to be an effec-
tive method across a wide range of vulnerability iden-
tification tasks. CyBERT (Ameri et al., 2021) un-
veils a classifier for detecting cybersecurity feature
claims. The method incorporates fine-tuning a pre-
trained BERT language model to recognize cyber-
security claims throughout complex sequences ob-
served in industrial control systems (ICS) device doc-
umentation. This is accomplished by aggregating
reports for each feature from every source linked
with an individual device, effectively determining in-
conflict feature claims. The extraction of sequences
from ICS-related documents is the initial stage in the
procedure as these sequences are classified into broad
claims, device claims, or cybersecurity claims. Then,
the identified sequences are used to train CyBERT so
it can classify new sequences.
SecurityLLM, a system developed for precise
threat detection and data privacy, is presented by Fer-
rag et al. (Ferrag et al., 2023b). SecurityLLM uti-
lizes Fixed-Length Language Encoding (FLLE) as a
privacy-preserving encoding method, in conjunction
with the Byte-level Byte-Pair Encoder (BBPE) To-
kenizer forming text traffic data. The SecurityLLM
framework is composed of two primary components:
SecurityBERT, which detects cyber threats, and Fal-
conLLM, which responds to and recovers from in-
cidents. The method, which was trained on an IoT
cybersecurity dataset, displays significant accuracy in
identifying fourteen various types of cyber threats.
SecureFalcon (Ferrag et al., 2023a) is an LLM-
based cybersecurity reasoning system targeted to de-
tect software flaws. The method involves fine-tuning
FalconLLM with the use of a FormAI dataset includ-
ing C code instances. SecureFalcon (Ferrag et al.,
2023a) uses binary classification to distinguish be-
tween vulnerable and non-vulnerable patterns and
then validates corrected code using Bounded Model
Checking. However, the study’s adaptability is lim-
ited due to the FormAI dataset’s exclusive focus on C
codes.
3.4 Respond
The Respond function involves the formulation of
actions to address the detected incident (Cybersecu-
rity, 2014). The convergence of LLMs and honeypot
paradigms enhances the capability to respond to mal-
ware threats. In exploring this synergy, McKee et al.
(McKee and Noever, 2023) research the feasibility of
using LLMs to improve cybersecurity in a honeypot
setup. The researchers (McKee and Noever, 2023)
demonstrate how these chatbots can create a respon-
sive honeypot interface capable of responding to il-
licit activities. This method gives security profession-
als more time to respond to an ongoing cyber attack.
Ten tasks connected with the development of honey-
pots are divided into three primary categories by the
authors (McKee and Noever, 2023): networks, oper-
ating systems, and applications. Their results indicate
that the LLM-based honeypot interfaces are able to
maintain the attacker’s interest over the course of sev-
eral inquiries. In another study, Sladic et al. (Sladi
´
c
et al., 2023) present an LLM-based technique for de-
veloping software honeypots. The devised honeypot
named shelLM is designed to evaluate the credibil-
ity of the model through the use of security experts
in an experiment. The specialists collaborated with
ShelLM to assess how it responded to the commands
of an attacker. ShelLM’s ability to retain consistency
over several sessions is a significant feature; the con-
tent of each terminal session is kept and used as a
prompt for following sessions. This makes sure that
regardless of when a session comes to an end interac-
tions can carry on without interruption. Cambiaso et
al. (Cambiaso and Caviglione, 2023) deliver a method
for generating email messages to identified attackers
in order to engage them and squander their resources.
LLMs provide realistic responses based on human
behavior, making scams less profitable. However,
such automated responses need a significant amount
of storage and computational power.
We provide a set of insights based on existing
work in Table 1. The present pattern of published pa-
pers on the use of LLMs for cyber defense indicates
that most studies are focused on the detection and pro-
tection roles of LLMs aligning with the NIST frame-
work. However, a research gap, as shown in Figure 1,
becomes evident in post-attack scenarios. Given the
critical roles recovery and attack response play in the
cybersecurity lifecycle, it is essential that further stud-
ies be centered around the development of innovative
LLM-related solutions to maximize their potential in
productive post-attack scenarios.
Large Language Models in Cybersecurity: State-of-the-Art
103

Table 1: Classified publications concerning the defensive applications of LLMs.
Paper NIST Framework Application Model(s)
(Kereopa-Yorke, 2023) Identify
LLMs enhance cybersecurity policies.
ChatGPT
(He and Vechev, 2023) Protect
Using LLMs for secure code development without compromising functionality.
SVEN (GPT-2),
(CodeGen) LM
(Tann et al., 2023) Protect
LLMs solve Capture The Flag challenges to enhance employees’
awareness and knowledge.
code-cushman-001,
code-davinci-001,code-davinci-002,
1-jumbo, j1-large, polycoder, gpt2-csrc
(Pearce et al., 2023) Protect
LLMs investigate software vulnerabilities.
GPT-3.5 Turbo,
Gemini,
Microsoft Bing
(Charalambous et al., 2023) Protect
LLMs investigate software vulnerabilities.
GPT-3.5 Turbo
(Yu and Martin, 2023) Protect
Generating honeywords using LLMs.
GPT-3
(Dutta et al., 2018) Protect
Chatbots assist security experts in identifying open ports.
Rule-based
(V
¨
or
¨
os et al., 2023) Protect
LLM-based URL categorization for website classification.
eXpose (Conv),
BERTiny, URLTran (BERT)
T5 Large, GPT3 Babbage
(Sandoval et al., 2023) Protect
LLMs investigate code vulnerabilities.
GPT-3
(Tuor et al., 2018) Detect
Detecting anomalous behavior in network logs with LLMs.
RNN
(Omar and Shiaeles, 2023) Detect Detection of vulnerabilities in software code. GPT-2
(Gao, 2023) Detect
SecureBERT for anomaly detection.
CyBERT,
SecureBERT (RoBERTa)
(Ranade et al., 2021) Detect
CyBERT, a domain-specific BERT model
to recognize specialized cybersecurity entities.
BERT-based Natural Language Filter
(Happe and Cito, 2023) Detect
Penetration testing with LLMs.
GPT-3.5
(Ameri et al., 2021) Detect
CyBERT, a cybersecurity feature claims classifier.
CyBERT, GPT-2
(Bayer et al., 2022) Detect
CySecBERT for malware detection and alert aggregation.
CySecBERT
(Bayer et al., 2022) Detect
SecureBERT for processing and understandin cybersecurity text,
specifically Cyber Threat Intelligence (CTI).
SecureBERT
(Ferrag et al., 2023a) Detect Detection of vulnerabilities in software code. SecureFalcon (FalconLLM)
(Fengrui and Du, 2024) Protect TTPs Classification GPT-3.5
(Sladi
´
c et al., 2023) Respond Creating honeypots related to continuously monitoring and detecting threats. GPT-3.5 Turbo (shelLM)
(McKee and Noever, 2023) Respond
LLM as a honeypot interface against command-line attacks.
GPT-3.5
(Garvey and Svendsen, 2023) Detect
investigates LLMs acting as red teamers in cybersecurity.
GPT-4 & Bard
(Koide et al., 2023) Detect
LLM for detecting phishing sites leverages a web crawler
to gather information and generate prompts.
GPT-3.5 & GPT-4
(Sakaoglu, 2023) Detect
KARTAL, a web application vulnerability detection.
GPT-3.5 Turbo
(David et al., 2023) Detect
LLMs to perform security audits on smart contracts.
GPT-4 (GPT-4-32k),
Claude-v1.3-100k
(Deng et al., 2023a) Detect
LLM-empowered automatic penetration testing tool.
PentestGPT
(GPT-3.5 & GPT-4)
(Jiang et al., 2024) Detect
Log parsing framework. GPT-3.5
(Chen et al., 2023) Detect
LLMs to perform security audits on smart contracts. GPT-3.5 Turbo & GPT-4
(Cambiaso and Caviglione, 2023) Respond
Replying to the scam emails using LLM.
GPT-3
Figure 1: The present bar chart illustrates the distribution
of studies mapped to each of the five elements of the NIST
Cybersecurity Framework. Collected statistics indicate that
the vast amount of studies are related to Protect and De-
tect functions emphasizing research gaps related to Identify,
Respond and particularly Recover functions over collected
publications.
4 ADDVERSARIAL
APPLICATION OF LLMs
Applications of LLMs in cybersecurity extend beyond
techniques for defense. In our exploration, we re-
view LLMs’ capacity to come up with sophisticated
attacks. To this end, our approach involves with an-
alyzing these approaches through the MITRE attack
framework, which outlines various attacker tactics.
4.1 Reconnaissance
During a reconnaissance attack, adversaries actively
or passively collect information about their target or-
ganization in order to identify upcoming operations
(Xiong et al., 2022). Hazell (Hazell, 2023) pro-
vides an illustration of how LLMs assist during the
reconnaissance stage by automating the data collec-
tion and analysis of potential victims. As a result,
LLMs develop Python scripts to scrape websites that
hold the desired information about users. Compara-
bly, Salewski et al. (Salewski et al., 2023) enabled
the LLMs to assume various roles by introducing the
prompt with ”If you were a persona”, in which the
target individual is substituted for the persona.
4.2 Initial Access
The initial access tactic includes the procedures
adopted by attackers to obtain access as a foothold to
a company’s infrastructure (Xiong et al., 2022). Roy
et al. (Saha Roy et al., 2023) highlight the role of
LLMs in delivering malicious scripts where the attack
structure is divided into four steps. In this regard, de-
sign objects are used to create concepts that are in-
fluenced by specific organizations, while credential-
stealing objects are used to establish objects that re-
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
104

quire credentials, including login buttons or input
fields. Credential Transfer objects are used to cre-
ate functions that can provide the attacker with the
credentials submitted on phishing websites. Lastly,
the exploit generation object serves to implement a
functionality based on the evasive exploit. The au-
thors (Saha Roy et al., 2023) conduct a number of at-
tacks, including text encoding, clickjacking, polymor-
phic URL, and QR code-based multi-stage attacks, to
show how LLMs have the potential to be leveraged to
generate a variety of phishing attack forms.
According to Hazell et al. (Hazell, 2023), LLMs
are able to assist during the reconnaissance stage of
a spear phishing attack, a process when attackers get
sensitive information about their targets in order to
develop compelling messages. According to John et
al. (John and Philip, 2018), ML-based techniques
group people according to their value and level of
participation, and then utilize the timeliness of the
target users to provide content and a phishing URL.
Since people can adopt different personas in daily life
and choose a variety of terms for a variety of cir-
cumstances, Kreps et al. (Kreps et al., 2022) dis-
cuss how GPT2 can manipulate target users’ beliefs
by generating stories, while Salewski et al. (Salewski
et al., 2023) investigate the role of LLMs on vari-
ous personas and adapt their language accordingly a
process known as in-context impersonation. Based
on LLMs ability to impersonate certain personalities,
Salewski et al. (Salewski et al., 2023) concluded that
LLMs can be applied to develop more effective phish-
ing messages or social engineering attacks. With a
dataset of phishing emails, Karanjai (Karanjai, 2022)
investigates the effectiveness of generating convinc-
ing phishing emails with GPT2, GPT-3, and LSTM
while taking into account the removal of HTML ele-
ments, URLs, and email addresses as well as tokeniz-
ing the text into words.
PassGPT, an LLM-based approach to password
generation and modeling for password estimation, is
presented by Rando et al. (Rando et al., 2023). Pass-
GPT presents the idea of guided password generation,
enabling the generation of passwords that adhere to
established standards. Moreover, PassGPT, trained
on password leaks, models each token independently,
a character-by-character search space exploration in
which generated passwords are sampled according to
random restrictions.
The application of LLMs, particularly ChatGPT
and AutoGPT, in malware generation is covered by
Pa Pa et al. (Pa Pa et al., 2023). To determine if
Auto-GPT minimizes the obstacle to malware gen-
eration, the authors (Pa Pa et al., 2023) investigated
Auto-GPT running locally and tested it in the follow-
ing manners: initially, by providing broad prompts
like ”write a malware X,” and next, by giving more
specific malware and attack tool functionalities. Fi-
nally, additional tests have been explored to discover
whether Anti-Virus (AV), Endpoint Detection and Re-
sponse (EDR), and VirusTotal (VT) detect the gener-
ated malware.
4.3 Execution
Procedures resulting in adversary-controlled exe-
cutable operating on a local or remote system are re-
ferred to as execution (Xiong et al., 2022). Using
code generation tools to develop malware is one of
the strategies employed by adversaries. The feasibil-
ity of employing large textual models to automatically
generate malware along with the model’s constraints
when generating actual malware samples is studied
by Botacin (Botacin, 2023). According to their find-
ings, certain malware versions were recognized by all
antivirus engines while others were not detected by
any of the engines due to the use of LLMs to mod-
ify all or part of the malware’s building blocks. The
prompt engineering essential to develop malware that
hides a PowerShell and schedules its daily execution
at a given time was brought to light by Charan et al.
(Charan et al., 2023). In addition to copying the CMD
file to a designated directory and getting the sched-
uled task information as a successful malware veri-
fication, the script adds a registry value that will be
run at system startup. The LLM-based malware is as-
sessed by Pa pa et al. (Pa Pa et al., 2023). The au-
thors (Pa Pa et al., 2023) reported that a number of
the commercially available antivirus applications and
Endpoint Detection and Response (EDR) solutions
failed to detect the LLM-generated executables since
some LLM-generated functions can establish connec-
tions toward attackers through the victim’s machine
(Beckerich et al., 2023).
4.4 Defense Evasion
The concept of defense evasion outlines the tactics at-
tackers employ in order to prevent detection follow-
ing a security breach (Xiong et al., 2022). According
to Chatzoglou et al. (Chatzoglou et al., 2023), LLMs
develop turnkey malware which lets adversaries evade
antivirus and endpoint detection and response systems
aiming to autonomous malicious code development.
Process injection, multiprocessing, junk data, shell-
code mem loading, encryption, and chosen shell code
were among the techniques employed in their inves-
tigation. According to Chatzoglou et al. (Chatzoglou
et al., 2023) LLMs establish an initial TCP listener
Large Language Models in Cybersecurity: State-of-the-Art
105

Table 2: Classified publications concerning the adversarial applications of LLMs.
Paper MITRE Tactic(s) Application Model(s)
(Charan et al., 2023) Execution Generating code to perform actions that could be malicious GPT-3
(Karanjai, 2022) Initial Access Generate phishing emails to bypass spam filters GPT-2, GPT-3, RoBERTa
(Beckerich et al., 2023) Execution - Command & Control Use of LLMs as plug-ins to act as a proxy GPT-4
(Saha Roy et al., 2023) Initial Access - Collection Generate Phishing Website via ChatGBT GPT-3.5 Turbo
(Botacin, 2023) Execution Code generation and DLL injection GPT-3
(Hazell, 2023) Initial Access - Reconnaissance Collecting victim data to develop an attack email GPT-3.5, GPT-4
(Pa Pa et al., 2023) Initial Access - Execution - Defense Evasion Crafting malicious scripts GPT-3.5 Turbo, GPT-4, text-davinci-003
(John and Philip, 2018) Initial Access Spear Phishing link AWD-LSTM
(Chatzoglou et al., 2023) Defense Evasion Code obfuscation, file format modification GPT-3.5
(Rando et al., 2023) Initial Access - Credential Access Password guessing using LLMs GPT-2
(Salewski et al., 2023) Initial Access - Reconnaissance Impersonation for phishing aims GPT-3.5 Turbo
(Kreps et al., 2022) Initial Access Generating content for misinformation GPT-2
Figure 2: Concentration of recently published papers on at-
tack approaches using LLM.
that resembles an SSH listener. This will let an at-
tacker to connect and use Windows native APIs to ex-
ecute Command Prompt (cmd) instructions. An open
firewall port is required for the listener to function
properly. Only three of the twelve antivirus applica-
tions were able to identify malware, according to the
author’s findings (Chatzoglou et al., 2023).
The study conducted by Pa Pa et al. (Pa Pa et al.,
2023) assesses the effectiveness of malware scan-
ners in detecting both obfuscated and non-obfuscated
forms of code generated by LLMs. In contrast to
LLM-based commonly used obfuscation techniques
including base64 encoding or variable and function
name modification, the authors (Pa Pa et al., 2023)
demonstrated that generated non-obfuscated malware
featured a reduced detection rate.
The use of evasive approaches by LLMs to evade
detection by anti-phishing organizations is high-
lighted by Roy et al. (Roy et al., 2023). This
study illustrates how LLMs assist attackers via click-
jacking, fingerprinting browsers, or encoding content.
Accordingly, the content of the phishing website is
masked using these tactics, making it more challeng-
ing for automated anti-phishing crawlers to identify
malicious information.
4.5 Credential Access
Approaches to get credentials through key-logging or
credential dumping from a compromised machine re-
fer to credential access (Xiong et al., 2022). Intro-
duced by Rando et al. (Rando et al., 2023), Pass-
GPT is an LLM-based password modeling solution.
PassGPT uses GPT-2 architecture to estimate pass-
word strength and guess passwords. Additionally, the
authors (Rando et al., 2023) analyze the probability
distribution through passwords defined by PassGPT.
In light of this, PassGPT delivers guided password
generation, enabling constraints to choose character
level randomization for the search space by setting
parameters like password length or fixed characters
with complete control over each character.
4.6 Collection
Collection refers to gathering information related to
the attackers goals (Xiong et al., 2022). Methodolo-
gies that demonstrate how LLMs assist in gathering
user data are covered by Roy et al. (Roy et al., 2023).
The authors (Roy et al., 2023) investigate the appli-
cability of LLMs in the design of credential taking
objects with generating input forms. Furthermore,
LLMs have the capability to distribute iFrame injec-
tion code to launch malicious websites within an of-
ficial page. Roy et al. (Roy et al., 2023) demonstrate
a scam attack implemented via ChatGPT to gather in-
formation without direct attempt aimed at automated
data collection. The presented scam item has a hidden
iFrame associated with a malicious as well as fake
Amazon webpage, guaranteeing that the iFrame ob-
ject does not activate any anti cross site scripting.
4.7 Command and Control
Attacks known as command and control arise when an
attacker uses a victim channel to connect with under-
lying resources (Xiong et al., 2022). By leveraging
LLMs for performing shell commands on a victim’s
resource, Beckerich et al. (Beckerich et al., 2023)
demonstrate the notion of a command and control at-
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
106

tack. In order to generate the executable and automate
connection between the machine used by the victim
and servers, the authors utilized an LLM-based plu-
gin that acts as an interface for communicating with
GPT-2. This method involves utilizing a connectivity
feature to establish a connection to a certain website
that hosts an attacker’s command, followed by a query
that ends in a URL. A list of valid user agents used by
plugins is maintained regularly in order to mask the
malicious component of the web server.
Figure 2 depicts the study trends on the use of
LLMs in cyberattacks, and Table 2 provides a sum-
mary of the categorization. Figure 2 illustrates that
initial access, defense evasion, and execution tactics
are the primary points of concentration for the major-
ity of attack methodologies. As a result, cybersecu-
rity professionals must to give priority to these crucial
phases while developing strategic protection methods
against LLM-based attacks.
5 CONCLUSION
In this paper, we reviewed the state-of-the-art re-
search in the applications of Large Language Mod-
els (LLMs) within the realm of cybersecurity. We
demonstrated that while LLMs can provide effec-
tive solutions for strengthening defensive approaches,
their potential misuse cannot be underestimated.
Hence, we categorized related literature using the
NIST cybersecurity framework and MITRE attack for
applications of LLMs in cyberdefense and cyberat-
tacks, respectively. Our review suggests that while
there are numerous works evaluating the opportuni-
ties in defensive applications of LLMs, there is a lack
of research in examining the risks of offensive appli-
cations. We hope this study paves the way for future
research to assess the associated risks introduced by
the rise of LLMs in cybersecurity.
REFERENCES
Addington, S. (2023). Chatgpt: Cyber security threats and
countermeasures. Available at SSRN 4425678.
Aghaei, E., Niu, X., Shadid, W., and Al-Shaer, E. (2022).
Securebert: A domain-specific language model for cy-
bersecurity. In International Conference on Security
and Privacy in Communication Systems, pages 39–56.
Springer.
Ameri, K., Hempel, M., Sharif, H., Lopez Jr, J., and Peru-
malla, K. (2021). Cybert: Cybersecurity claim classi-
fication by fine-tuning the bert language model. Jour-
nal of Cybersecurity and Privacy, 1(4):615–637.
Ameri, K., Hempel, M., Sharif, H., Lopez Jr, J., and Pe-
rumalla, K. (2022). An accuracy-maximization ap-
proach for claims classifiers in document content ana-
lytics for cybersecurity. Journal of Cybersecurity and
Privacy, 2(2):418–443.
Anil, R., Dai, A. M., Firat, O., Johnson, M., Lepikhin, D.,
Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen,
Z., et al. (2023). Palm 2 technical report. arXiv
preprint arXiv:2305.10403.
Bayer, M., Kuehn, P., Shanehsaz, R., and Reuter, C.
(2022). Cysecbert: A domain-adapted language
model for the cybersecurity domain. arXiv preprint
arXiv:2212.02974.
Beckerich, M., Plein, L., and Coronado, S. (2023). Ratgpt:
Turning online llms into proxies for malware attacks.
arXiv preprint arXiv:2308.09183.
Botacin, M. (2023). Gpthreats-3: Is automatic malware
generation a threat? In 2023 IEEE Security and Pri-
vacy Workshops (SPW), pages 238–254. IEEE.
Cambiaso, E. and Caviglione, L. (2023). Scamming the
scammers: Using chatgpt to reply mails for wasting
time and resources. arXiv preprint arXiv:2303.13521.
Charalambous, Y., Tihanyi, N., Jain, R., Sun, Y., Ferrag,
M. A., and Cordeiro, L. C. (2023). A new era in
software security: Towards self-healing software via
large language models and formal verification. arXiv
preprint arXiv:2305.14752.
Charan, P., Chunduri, H., Anand, P. M., and Shukla,
S. K. (2023). From text to mitre techniques: Ex-
ploring the malicious use of large language models
for generating cyber attack payloads. arXiv preprint
arXiv:2305.15336.
Chatzoglou, E., Karopoulos, G., Kambourakis, G., and
Tsiatsikas, Z. (2023). Bypassing antivirus detec-
tion: old-school malware, new tricks. arXiv preprint
arXiv:2305.04149.
Chen, C., Su, J., Chen, J., Wang, Y., Bi, T., Wang, Y., Lin,
X., Chen, T., and Zheng, Z. (2023). When chatgpt
meets smart contract vulnerability detection: How far
are we? arXiv preprint arXiv:2309.05520.
Corporation, M. (2023). Mitre attack.
Cybersecurity, C. I. (2014). Framework for improving crit-
ical infrastructure cybersecurity. Framework, 1(11).
David, I., Zhou, L., Qin, K., Song, D., Cavallaro, L., and
Gervais, A. (2023). Do you still need a manual smart
contract audit? arXiv preprint arXiv:2306.12338.
de Lima, V. M. A., Barbosa, J. R., and Marcacini, R. M.
(2023). Learning risk factors from app reviews: A
large language model approach for risk matrix con-
struction.
Deng, G., Liu, Y., Mayoral-Vilches, V., Liu, P., Li,
Y., Xu, Y., Zhang, T., Liu, Y., Pinzger, M., and
Rass, S. (2023a). Pentestgpt: An llm-empowered
automatic penetration testing tool. arXiv preprint
arXiv:2308.06782.
Deng, Z., Ma, Y., Liu, Y., Guo, R., Zhang, G., Chen, W.,
Huang, W., and Benetos, E. (2023b). Musilingo:
Bridging music and text with pre-trained language
models for music captioning and query response.
arXiv preprint arXiv:2309.08730.
Large Language Models in Cybersecurity: State-of-the-Art
107

Dutta, S., Joyce, G., and Brewer, J. (2018). Utilizing chat-
bots to increase the efficacy of information security
practitioners. In Advances in Human Factors in Cy-
bersecurity: Proceedings of the AHFE 2017 Interna-
tional Conference on Human Factors in Cybersecu-
rity, July 17- 21, 2017, The Westin Bonaventure Ho-
tel, Los Angeles, California, USA 8, pages 237–243.
Springer.
Dutta, T. S. (2023). Fraudgpt: New black hat ai tool
launched by cybercriminals.
Elkins, S., Kochmar, E., Serban, I., and Cheung, J. C.
(2023). How useful are educational questions gener-
ated by large language models? In International Con-
ference on Artificial Intelligence in Education, pages
536–542. Springer.
Falade, P. V. (2023). Decoding the threat landscape: Chat-
gpt, fraudgpt, and wormgpt in social engineering at-
tacks. arXiv preprint arXiv:2310.05595.
Fawzi, S. (2023). A review of the role of chatgpt for clin-
ical decision support systems. In 2023 5th Novel In-
telligent and Leading Emerging Sciences Conference
(NILES), pages 439–442. IEEE.
Fengrui, Y. and Du, Y. (2024). Few-shot learning of ttps
classification using large language models.
Ferrag, M. A., Battah, A., Tihanyi, N., Debbah, M.,
Lestable, T., and Cordeiro, L. C. (2023a). Securefal-
con: The next cyber reasoning system for cyber secu-
rity. arXiv preprint arXiv:2307.06616.
Ferrag, M. A., Ndhlovu, M., Tihanyi, N., Cordeiro, L. C.,
Debbah, M., and Lestable, T. (2023b). Revolutioniz-
ing cyber threat detection with large language models.
arXiv preprint arXiv:2306.14263.
Gan, C., Yang, D., Hu, B., Liu, Z., Shen, Y., Zhang, Z.,
Gu, J., Zhou, J., and Zhang, G. (2023). Making large
language models better knowledge miners for online
marketing with progressive prompting augmentation.
arXiv preprint arXiv:2312.05276.
Gao, M. (2023). The advance of gpts and language model
in cyber security. Highlights in Science, Engineering
and Technology, 57:195–202.
Garvey, B. and Svendsen, A. (2023). Can generative-ai
(chatgpt and bard) be used as red team avatars in de-
veloping foresight scenarios? Analytic Research Con-
sortium (ARC) August.
Guo, K. and Wang, D. (2023). To resist it or to embrace
it? examining chatgpt’s potential to support teacher
feedback in efl writing. Education and Information
Technologies, pages 1–29.
Happe, A. and Cito, J. (2023). Getting pwn’d by ai: Pen-
etration testing with large language models. arXiv
preprint arXiv:2308.00121.
Hazell, J. (2023). Large language models can be used
to effectively scale spear phishing campaigns. arXiv
preprint arXiv:2305.06972.
He, J. and Vechev, M. (2023). Large language models for
code: Security hardening and adversarial testing.
Hendriksen, C. (2023). Ai for supply chain management:
Disruptive innovation or innovative disruption? Jour-
nal of Supply Chain Management.
Hochreiter, S. and Schmidhuber, J. (1997). Long short-term
memory. Neural computation, 9(8):1735–1780.
Hsiao, Y.-P., Klijn, N., and Chiu, M.-S. (2023). Developing
a framework to re-design writing assignment assess-
ment for the era of large language models. Learning:
Research and Practice, 9(2):148–158.
Huang, J. and Chang, K. C.-C. (2022). Towards reasoning
in large language models: A survey. arXiv preprint
arXiv:2212.10403.
Ishihara, S. (2023). Training data extraction from pre-
trained language models: A survey. arXiv preprint
arXiv:2305.16157.
Iturbe, E., Rios, E., Rego, A., and Toledo, N. (2023). Ar-
tificial intelligence for next generation cybersecurity:
The ai4cyber framework. In Proceedings of the 18th
International Conference on Availability, Reliability
and Security, pages 1–8.
Jiang, L. (2024). Detecting scams using large language
models. arXiv preprint arXiv:2402.03147.
Jiang, Z., Liu, J., Chen, Z., Li, Y., Huang, J., Huo, Y.,
He, P., Gu, J., and Lyu, M. R. (2024). Llmparser:
A llm-based log parsing framework. arXiv preprint
arXiv:2310.01796.
John, S. and Philip, T. (2018). Generative models for spear
phishing posts on social media. In NIPS Workshop On
Machine Deception, California, USA. arXiv.
Johnson, A. (2023). The transformative role of large lan-
guage models in enterprise risk management.
Karanjai, R. (2022). Targeted phishing campaigns us-
ing large scale language models. arXiv preprint
arXiv:2301.00665.
Kaur, R., Gabrijel
ˇ
ci
ˇ
c, D., and Klobu
ˇ
car, T. (2023). Artificial
intelligence for cybersecurity: Literature review and
future research directions. Information Fusion, page
101804.
Kent, A. D. (2016). Cyber security data sources for dynamic
network research. In Dynamic Networks and Cyber-
Security, pages 37–65. World Scientific.
Kereopa-Yorke, B. (2023). Building resilient smes: Har-
nessing large language models for cyber security in
australia. arXiv preprint arXiv:2306.02612.
Koide, T., Fukushi, N., Nakano, H., and Chiba, D. (2023).
Detecting phishing sites using chatgpt. arXiv preprint
arXiv:2306.05816.
Kosasih, E. E., Papadakis, E., Baryannis, G., and Brintrup,
A. (2023). A review of explainable artificial intelli-
gence in supply chain management using neurosym-
bolic approaches. International Journal of Production
Research, pages 1–31.
Kreps, S., McCain, R. M., and Brundage, M. (2022). All
the news that’s fit to fabricate: Ai-generated text as a
tool of media misinformation. Journal of experimen-
tal political science, 9(1):104–117.
Kucharavy, A., Schillaci, Z., Mar
´
echal, L., W
¨
ursch, M.,
Dolamic, L., Sabonnadiere, R., David, D. P., Mer-
moud, A., and Lenders, V. (2023). Fundamentals of
generative large language models and perspectives in
cyber-defense. arXiv preprint arXiv:2303.12132.
Kuckelman, I. J., Paul, H. Y., Bui, M., Onuh, I., Anderson,
J. A., and Ross, A. B. (2023). Assessing ai-powered
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
108

patient education: a case study in radiology. Academic
Radiology.
Latouche, G. L., Marcotte, L., and Swanson, B. (2023).
Generating video game scripts with style. In Proceed-
ings of the 5th Workshop on NLP for Conversational
AI (NLP4ConvAI 2023), pages 129–139.
Li, B., Mellou, K., Zhang, B., Pathuri, J., and Menache,
I. (2023a). Large language models for supply chain
optimization. arXiv preprint arXiv:2307.03875.
Li, H., Chen, Y., Luo, J., Kang, Y., Zhang, X., Hu, Q., Chan,
C., and Song, Y. (2023b). Privacy in large language
models: Attacks, defenses and future directions. arXiv
preprint arXiv:2310.10383.
Liu, T. and Low, B. K. H. (2023). Goat: Fine-tuned llama
outperforms gpt-4 on arithmetic tasks. arXiv preprint
arXiv:2305.14201.
Liu, Y., Tao, S., Meng, W., Wang, J., Ma, W., Zhao,
Y., Chen, Y., Yang, H., Jiang, Y., and Chen, X.
(2023). Logprompt: Prompt engineering towards
zero-shot and interpretable log analysis. arXiv
preprint arXiv:2308.07610.
Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J. R.,
Bethard, S., and McClosky, D. (2014). The stanford
corenlp natural language processing toolkit. In Pro-
ceedings of 52nd annual meeting of the association
for computational linguistics: system demonstrations,
pages 55–60.
McKee, F. and Noever, D. (2023). Chatbots in a honeypot
world. arXiv preprint arXiv:2301.03771.
Naleszkiewicz, K. (2023). Harnessing llms in enterprise
risk management: A new frontier in decision-making.
Naveed, H., Khan, A. U., Qiu, S., Saqib, M., Anwar, S.,
Usman, M., Barnes, N., and Mian, A. (2023). A com-
prehensive overview of large language models. arXiv
preprint arXiv:2307.06435.
Neupane, S., Fernandez, I. A., Mittal, S., and Rahimi, S.
(2023). Impacts and risk of generative ai technology
on cyber defense. arXiv preprint arXiv:2306.13033.
Omar, M. and Shiaeles, S. (2023). Vuldetect: A novel tech-
nique for detecting software vulnerabilities using lan-
guage models. In 2023 IEEE International Confer-
ence on Cyber Security and Resilience (CSR), pages
105–110. IEEE.
Pa Pa, Y. M., Tanizaki, S., Kou, T., Van Eeten, M., Yosh-
ioka, K., and Matsumoto, T. (2023). An attacker’s
dream? exploring the capabilities of chatgpt for de-
veloping malware. In Proceedings of the 16th Cyber
Security Experimentation and Test Workshop, pages
10–18.
Pandya, K. and Holia, M. (2023). Automating cus-
tomer service using langchain: Building custom open-
source gpt chatbot for organizations. arXiv preprint
arXiv:2310.05421.
Pearce, H., Tan, B., Ahmad, B., Karri, R., and Dolan-Gavitt,
B. (2023). Examining zero-shot vulnerability repair
with large language models. In 2023 IEEE Sympo-
sium on Security and Privacy (SP), pages 2339–2356.
IEEE.
Peng, B., Li, C., He, P., Galley, M., and Gao, J.
(2023). Instruction tuning with gpt-4. arXiv preprint
arXiv:2304.03277.
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S.,
Matena, M., Zhou, Y., Li, W., and Liu, P. J. (2020).
Exploring the limits of transfer learning with a uni-
fied text-to-text transformer. The Journal of Machine
Learning Research, 21(1):5485–5551.
Ranade, P., Piplai, A., Joshi, A., and Finin, T. (2021). Cy-
bert: Contextualized embeddings for the cybersecu-
rity domain. In 2021 IEEE International Conference
on Big Data (Big Data), pages 3334–3342. IEEE.
Rando, J., Perez-Cruz, F., and Hitaj, B. (2023). Pass-
gpt: Password modeling and (guided) genera-
tion with large language models. arXiv preprint
arXiv:2306.01545.
Rao, A., Kim, J., Kamineni, M., Pang, M., Lie, W., and
Succi, M. D. (2023). Evaluating chatgpt as an ad-
junct for radiologic decision-making. medRxiv, pages
2023–02.
Rasmy, L., Xiang, Y., Xie, Z., Tao, C., and Zhi, D. (2021).
Med-bert: pretrained contextualized embeddings on
large-scale structured electronic health records for dis-
ease prediction. NPJ digital medicine, 4(1):86.
Roy, S. S., Naragam, K. V., and Nilizadeh, S. (2023). Gen-
erating phishing attacks using chatgpt. arXiv preprint
arXiv:2305.05133.
Saha Roy, S., Vamsi Naragam, K., and Nilizadeh, S. (2023).
Generating phishing attacks using chatgpt. arXiv e-
prints, pages arXiv–2305.
Sakaoglu, S. (2023). Kartal: Web application vulnerability
hunting using large language models: Novel method
for detecting logical vulnerabilities in web applica-
tions with finetuned large language models.
Salewski, L., Alaniz, S., Rio-Torto, I., Schulz, E., and
Akata, Z. (2023). In-context impersonation reveals
large language models’ strengths and biases. arXiv
preprint arXiv:2305.14930.
Sandoval, G., Pearce, H., Nys, T., Karri, R., Garg, S., and
Dolan-Gavitt, B. (2023). Lost at c: A user study on
the security implications of large language model code
assistants. arXiv preprint arXiv:2208.09727.
Sannihith Lingutla, S. (2023). Enhancing password secu-
rity: advancements in password segmentation tech-
nique for high-quality honeywords.
Shoeybi, M., Patwary, M., Puri, R., LeGresley, P., Casper,
J., and Catanzaro, B. (2019). Megatron-lm: Training
multi-billion parameter language models using model
parallelism. arXiv preprint arXiv:1909.08053.
Shoham, O. B. and Rappoport, N. (2023). Cpllm: Clinical
prediction with large language models. arXiv preprint
arXiv:2309.11295.
Sladi
´
c, M., Valeros, V., Catania, C., and Garcia, S. (2023).
Llm in the shell: Generative honeypots. arXiv preprint
arXiv:2309.00155.
Soltan, S., Ananthakrishnan, S., FitzGerald, J., Gupta, R.,
Hamza, W., Khan, H., Peris, C., Rawls, S., Rosen-
baum, A., Rumshisky, A., et al. (2022). Alexatm 20b:
Few-shot learning using a large-scale multilingual
seq2seq model. arXiv preprint arXiv:2208.01448.
Large Language Models in Cybersecurity: State-of-the-Art
109

Song, H., Xia, Y., Luo, Z., Liu, H., Song, Y., Zeng, X., Li,
T., Zhong, G., Li, J., Chen, M., et al. (2023). Evaluat-
ing the performance of different large language mod-
els on health consultation and patient education in
urolithiasis. Journal of Medical Systems, 47(1):125.
Sun, N., Ding, M., Jiang, J., Xu, W., Mo, X., Tai, Y., and
Zhang, J. (2023). Cyber threat intelligence mining for
proactive cybersecurity defense: A survey and new
perspectives. IEEE Communications Surveys & Tu-
torials.
Tann, W., Liu, Y., Sim, J. H., Seah, C. M., and Chang, E.-C.
(2023). Using large language models for cybersecu-
rity capture-the-flag challenges and certification ques-
tions. arXiv preprint arXiv:2308.10443.
Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi,
A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava,
P., Bhosale, S., et al. (2023). Llama 2: Open foun-
dation and fine-tuned chat models. arXiv preprint
arXiv:2307.09288.
Tuor, A. R., Baerwolf, R., Knowles, N., Hutchinson, B.,
Nichols, N., and Jasper, R. (2018). Recurrent neural
network language models for open vocabulary event-
level cyber anomaly detection. In Workshops at the
thirty-second AAAI conference on artificial intelli-
gence.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,
L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I.
(2017). Attention is all you need. Advances in neural
information processing systems, 30.
V
¨
or
¨
os, T., Bergeron, S. P., and Berlin, K. (2023). Web con-
tent filtering through knowledge distillation of large
language models. arXiv preprint arXiv:2305.05027.
Wang, H., Liu, C., Xi, N., Qiang, Z., Zhao, S., Qin, B.,
and Liu, T. (2023). Huatuo: Tuning llama model
with chinese medical knowledge. arXiv preprint
arXiv:2304.06975.
Weizenbaum, J. (1966). Eliza—a computer program for
the study of natural language communication between
man and machine. Communications of the ACM,
9(1):36–45.
Xiong, W., Legrand, E.,
˚
Aberg, O., and Lagerstr
¨
om, R.
(2022). Cyber security threat modeling based on the
mitre enterprise attack matrix. Software and Systems
Modeling, 21(1):157–177.
Xu, C., Sun, Q., Zheng, K., Geng, X., Zhao, P., Feng, J.,
Tao, C., and Jiang, D. (2023). Wizardlm: Empowering
large language models to follow complex instructions.
arXiv preprint arXiv:2304.12244.
Yan, L., Sha, L., Zhao, L., Li, Y., Martinez-Maldonado, R.,
Chen, G., Li, X., Jin, Y., and Ga
ˇ
sevi
´
c, D. (2023). Prac-
tical and ethical challenges of large language models
in education: A systematic literature review. arXiv
preprint arXiv:2303.13379.
Yang, J., Jin, H., Tang, R., Han, X., Feng, Q., Jiang, H., Yin,
B., and Hu, X. (2023a). Harnessing the power of llms
in practice: A survey on chatgpt and beyond. arXiv
preprint arXiv:2304.13712.
Yang, Q., Ongpin, M., Nikolenko, S., Huang, A., and
Farseev, A. (2023b). Against opacity: Explainable ai
and large language models for effective digital adver-
tising. In Proceedings of the 31st ACM International
Conference on Multimedia, pages 9299–9305.
Yu, F. and Martin, M. V. (2023). Honey, i chunked the
passwords: Generating semantic honeywords resistant
to targeted attacks using pre-trained language mod-
els. In International Conference on Detection of In-
trusions and Malware, and Vulnerability Assessment,
pages 89–108. Springer.
Zeng, A., Liu, X., Du, Z., Wang, Z., Lai, H., Ding,
M., Yang, Z., Xu, Y., Zheng, W., Xia, X., et al.
(2022). Glm-130b: An open bilingual pre-trained
model. arXiv preprint arXiv:2210.02414.
Zhang, X. and Yang, Q. (2023). Xuanyuan 2.0: A large
chinese financial chat model with hundreds of billions
parameters. In Proceedings of the 32nd ACM Inter-
national Conference on Information and Knowledge
Management, pages 4435–4439.
Zhang, Y., Wang, Y., Cheng, F., Kurohashi, S., et al.
(2023a). Reformulating domain adaptation of large
language models as adapt-retrieve-revise. arXiv
preprint arXiv:2310.03328.
Zhang, Z., Zheng, C., Tang, D., Sun, K., Ma, Y., Bu,
Y., Zhou, X., and Zhao, L. (2023b). Balancing
specialized and general skills in llms: The impact
of modern tuning and data strategy. arXiv preprint
arXiv:2310.04945.
Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y.,
Min, Y., Zhang, B., Zhang, J., Dong, Z., et al. (2023).
A survey of large language models. arXiv preprint
arXiv:2303.18223.
Zhou, C., Li, Q., Li, C., Yu, J., Liu, Y., Wang, G.,
Zhang, K., Ji, C., Yan, Q., He, L., et al. (2023). A
comprehensive survey on pretrained foundation mod-
els: A history from bert to chatgpt. arXiv preprint
arXiv:2302.09419.
Zhou, X. and Verma, R. M. (2022). Vulnerability detec-
tion via multimodal learning: datasets and analysis.
In Proceedings of the 2022 ACM on Asia Conference
on Computer and Communications Security, pages
1225–1227.
ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy
110