Large Language Models in Cybersecurity: State-of-the-Art

1 INTRODUCTION

• LLMs’ ability to process complex natural-language patterns lets security experts explore attack vectors involving textual data.The paper situates this capability amid broader LLM adoption across fields including education, entertainment, business, and healthcare.Source: §1 · p. 1
• LLMs can support cybersecurity defense through log analysis, anomaly detection, vulnerability identification, and automated fixes.These applications aim to identify emerging vulnerabilities and improve organizations’ security posture.Source: §1 · p. 1
• LLMs can also be misused to execute malware, evade defenses, obtain credentials, and generate personalized phishing messages.WormGPT is described as automating personalized phishing emails, while FraudGPT generated content to persuade users to click a link.Source: §1 · p. 1
• The paper frames LLM-enabled offensive approaches with the MITRE attack framework and defensive strategies with the NIST cybersecurity framework.It also surveys major LLM functionalities and potential cybersecurity applications.Source: §1 · pp. 1–2

2 BACKGROUND

• LLMs are neural networks trained on textual data to perform language-related tasks.The background traces NLP developments from early chatbots through long short-term memory systems and Stanford CoreNLP to transformer architecture.Source: §2 · p. 2
• Transformer advances enabled diverse LLMs, while fine-tuning adapts general-purpose models to industry-specific criteria.The background also notes deployment challenges such as high costs, hardware requirements, and slow training, alongside retrieval, parameter tuning, and distillation approaches.Source: §2 · p. 2
• Unlike prior surveys of LLM architecture or safety, this review focuses on LLM applications in cyberdefense and cyberattack.The cited prior literature includes surveys addressing LLM safety and architecture.Source: §2 · p. 2

3 DEFENSIVE APPLICATIONS OF LLMS

• The review classifies defensive LLM approaches using NIST’s functions: Identify, Protect, Detect, Respond, and Recover.It uses the framework to organize studies according to their role in managing cybersecurity threats and incidents.Source: §3 · p. 2

3.1 Identify

• LLMs support risk identification by analyzing diverse sources, classifying application-review risks, and synthesizing organizational data into risk profiles and plans.The reviewed work also describes LLM-generated risk assessments, contingency plans, and reporting.Source: §3.1 · p. 2

3.2 Protect

• Protect applications include web-content filtering, GPT-3 honeyword generation, and LLM-assisted cybersecurity education and team collaboration.The honeyword approach depends on preserving password components, or “chunks”; models also face limitations on conceptual queries.Source: §3.2 · pp. 2–3
• LLM-assisted vulnerability repair uses Bounded Model Checking before and after code generation, but real-world fixing remains sensitive to defect definition and prompt phrasing.The reviewed work reports that simulated vulnerabilities were fully fixed, while real-world scenarios presented performance challenges.Source: §3.2 · p. 3

3.3 Detect

• Other detection work spans CTI models, penetration-testing support, phishing-site analysis, and smart-contract auditing, where a separate study reported frequent false positives.The smart-contract studies include analysis of 52 compromised DeFi protocols and evaluation on 10,000 contracts across nine vulnerability types.Source: §3.3 · pp. 3–4
• Detection approaches include adaptive broken-access-control analysis, log anomaly scoring, and SecureFalcon code-flaw detection, whose adaptability is limited by its C-only dataset.Log analysis is described as real-time and as reducing false positives without significantly affecting unusual-behavior detection.Source: §3.3 · pp. 3–4

3.4 Respond

• Respond approaches include honeypot interfaces that maintain attackers’ interest and automated email engagement, though the latter requires substantial storage and computation.ShelLM also retains terminal-session content as context for later sessions.Source: §3.4 · pp. 4–5
• Most reviewed work addresses Protect and Detect, while Identify, Respond, and especially Recover have comparatively sparse coverage.The authors identify post-attack response and recovery as research gaps.Source: §3.4 · p. 5, Figure 1 · p. 5

4.1 Reconnaissance

• LLM-enabled reconnaissance includes scraping target information and prompting models to adopt target personas.These approaches are described as supporting collection and analysis of potential victims.Source: §4.1 · p. 5

4.2 Initial Access

• Studies cover LLM-generated phishing forms, persona-adapted messages, guided password modeling, and malware-generation tests.The malware tests included broad and specific prompts and checks by AV, EDR, and VirusTotal.Source: §4.2 · pp. 5–6

4.3 Execution

• LLM-modified malware samples varied in antivirus detection, and reported LLM-generated executables sometimes evaded commercial antivirus and EDR tools.A described script hid PowerShell and scheduled daily execution; it also added a registry value to run at startup.Source: §4.3 · p. 6

4.4 Defense Evasion

• Only three of twelve antivirus applications identified the turnkey malware studied, and generated non-obfuscated malware had a reduced detection rate.The review also describes process injection, encryption, and other evasion techniques, plus phishing-site tactics that mask content from crawlers.Source: §4.4 · pp. 6–7

4.5 Credential Access

• PassGPT uses GPT-2 to estimate password strength and guess passwords, with guided generation allowing constraints such as length and fixed characters.Source: §4.5 · p. 7

4.6 Collection

• LLMs can generate credential-taking input forms and iFrame injection code; one described ChatGPT scam used a hidden iFrame with a fake Amazon page.Source: §4.6 · p. 7

4.7 Command and Control

• An LLM-based plugin interfacing with GPT-2 generated executables and automated connections between victim machines and servers for command and control.The described method used a website-hosted attacker command and maintained valid user-agent lists to mask the server component.Source: §4.7 · p. 7
• Initial access, defense evasion, and execution are the primary concentration points among reviewed LLM-based attack methodologies.The paper says cybersecurity professionals should prioritize these phases when developing protection methods.Source: §4.7 · p. 7

• The review categorizes LLM applications in cyberdefense with the NIST cybersecurity framework and cyberattacks with MITRE.Source: §5 · p. 8
• The authors conclude that LLMs can strengthen defense, while their potential misuse cannot be underestimated.Source: §5 · p. 8
• The review finds numerous studies of defensive opportunities but a lack of research examining offensive risks, and calls for further risk assessment.Source: §5 · p. 8