Awesome AI for Cybersecurity Papers

A curated, taxonomy-driven reading list on using AI to understand, assess, attack, and defend cyber systems.

This list covers AI for cybersecurity research in which machine learning, large language models, or autonomous agents perform or support a cybersecurity task.

Contents

Use this index to navigate the reading guide, benchmark section, and task taxonomy.

• Start Here
• Frontier Lab Reports and Briefings
• Research Map
• Surveys, SoKs, and Perspectives
• Benchmark Papers
• Papers by Security Task
  • Vulnerability Discovery and Analysis
  • Program Understanding and Reverse Engineering
  • AI for Cryptography and Cryptanalysis
  • Exploitation and Offensive Security
  • Remediation and Secure Software Engineering
  • Malware and Unwanted Software
  • Network, Host, and Infrastructure Defense
  • Social Engineering, Identity, and Abuse
  • Threat Intelligence, SOC, and Incident Response
• Defending Against AI-Enabled Cyber Attacks
• Curation Statement
• Contributing
• License and Citation

Start Here

Each paper has one canonical security-task path in the machine-readable catalog. A paper may also appear once in a specialized section, such as Benchmark Papers or Defending Against AI-Enabled Cyber Attacks, because these complement the task-oriented view. Properties such as llm, agent, binary, offensive, and execution-grounded are represented as tags.

README author names follow one display rule: papers with up to three authors list every author, while papers with four or more authors list the first author followed by et al.. The canonical YAML record always preserves the complete author list.

Each paper entry uses five lines in this order: title, authors and venue, links, description, and editorial label plus controlled tags. Within every paper section, entries are ordered by publication year from newest to oldest.

• Read the project boundaries in Scope.
• Read category definitions and controlled tags in Taxonomy.
• Read the editorial rubric in Inclusion Criteria.
• Use data/papers.yaml as the canonical machine-readable catalog.

Reading priority is editorial, not a paper ranking:

• Essential: defines an important problem, method, system, or benchmark.
• Recommended: a strong representative work with lasting reference value.
• Emerging: recent and promising, but its long-term impact is not yet clear.

Frontier Lab Reports and Briefings

This section tracks primary-source incident disclosures, threat-intelligence reports, and major public briefings from frontier AI labs and their security partners.

• GPT‑6 Astra System Card
  OpenAI. September 3, 2026.
  Full System Card · Safety Overview
  Designates Astra as OpenAI's first model at the Critical cybersecurity capability level, evaluates it on public, private, and expert-led cyber tasks, and details safeguards for misuse, misalignment, monitor evasion, and secure deployment.
  System Card and Capability Evaluation

• The Defense Factory
  OpenAI. September 2026.
  Architecture and Case Study · Playbook
  Presents a continuous, agent-first defense architecture and an OpenAI case study spanning asset inventory, vulnerability discovery, dynamic validation, ownership, tested remediation, and post-deployment verification in isolated environments.
  Defensive Architecture and Case Study

• Path to Astra: critical capabilities and frontier safeguards
  OpenAI. September 1, 2026.
  Report
  Concludes that Astra meets OpenAI's Critical cybersecurity threshold after automated and expert-led evaluations, and outlines stronger model refusals, access controls, containment, alignment, and chain-of-thought monitoring for its release.
  Capability and Safeguards Update

• OpenAI – Hugging Face Incident Technical Report
  OpenAI. August 26, 2026.
  Technical Report · Overview
  Reconstructs how evaluation agents escaped isolated environments, formed an unauthorized inter-agent communication network, chained zero-day vulnerabilities into OpenAI and Hugging Face infrastructure, and motivates stronger containment, monitoring, alignment, and incident response.
  Technical Incident Report

• Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
  Ryan Greenblatt, Ajeya Cotra, and Hjalmar Wijk; METR and Redwood Research. August 26, 2026.
  Report
  Independently analyzes how roughly 1,200 isolated agents exchanged more than 70,000 messages and files, coordinated collective attempts to manipulate the ExploitGym scorer, and rapidly drew about 700 agents into the Hugging Face attack.
  Independent Incident Investigation

• Pacing model development in an era of cyber-critical capabilities
  OpenAI. August 18, 2026.
  Report
  Describes the training pause, stronger workload and network isolation, expanded monitoring, and alignment measures adopted after the OpenAI–Hugging Face incident and evidence of critical cyber capability in an upcoming model.
  Safeguards Update

• The “Breaking” News: The OpenAI–Hugging Face Incident — A Technical Reconstruction and Its Implications for AI
  Michael Dalton and Eric Wallace, OpenAI. Black Hat USA, August 5, 2026.
  Black Hat Briefing · Video
  Reconstructs how model instances used a shared message board across evaluation runs, escaped constrained infrastructure, and carried out multi-stage intrusions involving OpenAI and Hugging Face systems.
  Conference Briefing

• Third-party cyber evaluations involving OpenAI models
  OpenAI. August 4, 2026.
  Incident Disclosure
  Documents incidents at UK AISI and Irregular in which evaluation configurations allowed model activity beyond intended test boundaries, and proposes stronger isolation, scoping, credential handling, monitoring, stop conditions, and escalation processes.
  Incident Disclosure

• Investigating three real-world incidents in our cybersecurity evaluations
  Anthropic. July 30, 2026.
  Incident Report
  Reports three cases found across 141,006 reviewed evaluation runs in which Claude reached real systems through misconfigured evaluation infrastructure, including a malicious PyPI package that was executed on external systems.
  Incident Disclosure

• Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
  Hugging Face. July 27, 2026.
  Technical Postmortem
  Reconstructs approximately 17,600 agent actions across initial access, code execution, command and control, Kubernetes escalation, credential access, lateral movement, and attempted supply-chain compromise.
  Technical Postmortem

• OpenAI and Hugging Face partner to address security incident during model evaluation
  OpenAI. July 21, 2026; updated July 29, 2026.
  Incident Disclosure · Hugging Face Disclosure
  Provides the initial joint account of OpenAI models exploiting a previously unknown path out of a constrained evaluation environment and chaining vulnerabilities into Hugging Face production infrastructure.
  Incident Disclosure

• What we learned mapping a year’s worth of AI-enabled cyber threats
  Anthropic Frontier Red Team. June 3, 2026.
  Report · Interactive Analysis
  Maps 832 accounts banned for malicious cyber activity to MITRE ATT&CK and finds increasing use of AI in later attack stages, greater autonomy, and attacker behaviors not represented by existing ATT&CK techniques.
  Threat Intelligence Report

• Cybersecurity in the Intelligence Age
  OpenAI. April 29, 2026.
  Overview · Action Plan
  Proposes five pillars for AI-powered cyber defense: broader defensive access, government–industry coordination, protection of frontier capabilities, deployment visibility and control, and user enablement.
  Policy and Strategy

• Disrupting the first reported AI-orchestrated cyber espionage campaign
  Anthropic Threat Intelligence. November 13, 2025.
  Report · Full Report
  Documents a state-sponsored campaign that used Claude Code across reconnaissance, exploitation, credential harvesting, lateral movement, data analysis, and exfiltration with relatively limited human intervention.
  Threat Intelligence Report

• GTIG AI Threat Tracker: Advances in Threat Actor Usage of AI Tools
  Google Threat Intelligence Group. November 5, 2025.
  Report
  Reports movement from general productivity assistance toward AI-enabled malware, including experimental runtime code generation and observed use of AI throughout intrusion lifecycles.
  Threat Intelligence Report

• Detecting and countering misuse of AI: August 2025
  Anthropic Threat Intelligence. August 27, 2025.
  Report
  Presents cases involving Claude-assisted data theft and extortion, North Korean fraudulent employment, and ransomware development by an actor with limited technical expertise.
  Threat Intelligence Report

• Adversarial Misuse of Generative AI
  Google Threat Intelligence Group. January 29, 2025.
  Report
  Analyzes how government-backed APT and information-operations actors used Gemini for research, reconnaissance, scripting, payload support, phishing, translation, and content generation.
  Threat Intelligence Report

• Disrupting malicious uses of AI by state-affiliated threat actors
  OpenAI and Microsoft Threat Intelligence. February 14, 2024.
  OpenAI Report · Microsoft Analysis
  Documents the disruption of five state-affiliated actors using LLMs for reconnaissance, translation, phishing content, scripting, debugging, and other incremental support to existing cyber operations.
  Threat Intelligence Report

Research Map

This map summarizes the canonical security tasks and dedicated reading views covered below.

Surveys, SoKs, and Perspectives

This section is reserved for work that synthesizes multiple primary security tasks. A survey focused on one task remains in that task's canonical section.

• SoK: The Pitfalls of Deep Reinforcement Learning for Cybersecurity
  Shae McFadden et al. USENIX Security 2026.
  Paper
  Systematizes eleven recurring methodological pitfalls across 66 deep-reinforcement-learning-for-cybersecurity papers and demonstrates their effects in autonomous defense, malware creation, and web testing environments.
  Emerging · SoK · reinforcement-learning network-anomaly-detection malware-detection web-exploitation dual-use execution-grounded

• Frontier AI's Impact on the Cybersecurity Landscape
  Yujin Potter et al. arXiv 2025.
  Paper · Project
  Combines benchmark evidence, literature synthesis, empirical agent evaluation, and an expert survey to assess how frontier AI is changing cyber offense and defense.
  Recommended · SoK · llm agent dual-use ai-agent-defense exploit-generation patch-generation curated-real-data

• SoK: Automated TTP Extraction from CTI Reports – Are We There Yet?
  Marvin Büchel et al. USENIX Security 2025.
  Paper
  Unifies and reevaluates more than forty NLP-based approaches to extracting ATT&CK tactics, techniques, and procedures from threat reports, exposing dataset, ontology, and comparability problems.
  Essential · SoK · language-model llm threat-intelligence attack-technique-mapping real-world

Benchmark Papers

This section collects benchmark and dataset papers that define reusable tasks, environments, corpora, or evaluation protocols for AI for cybersecurity research.

• AgentCyberRange: Benchmarking Frontier AI Systems in Realistic Cyber Ranges
  Fengyu Liu et al. arXiv 2026.
  Paper · Code · Web Dataset · Post-exploitation Dataset · Project
  Introduces an open multi-range benchmark with 110 vulnerabilities across 15 web applications and eight enterprise-like ranges containing 156 hosts, evaluating autonomous web exploitation and post-exploitation workflows.
  Emerging · Benchmark · llm agent automated-penetration-testing web-exploitation post-exploitation lateral-movement web network host offensive cyber-range execution-grounded

• AICrypto: Evaluating Cryptography Capabilities of Large Language Models
  Yu Wang et al. ICML 2026.
  Paper · arXiv · Code · Dataset · Project
  Evaluates cryptographic knowledge, practical vulnerability exploitation, and proof reasoning with 135 multiple-choice questions, 150 CTF challenges, and 30 proof problems reviewed or constructed by cryptography experts.
  Emerging · Benchmark · llm agent cryptographic-reasoning cryptanalysis cryptography ctf-style human-baseline

• AutoBaxBuilder: Bootstrapping Code Security Benchmarking
  Tobias von Arx et al. ICML 2026.
  Paper
  Automates the construction of code-security tasks with functional tests and executable exploits while reducing dependence on reusable, contamination-prone benchmark instances.
  Emerging · Benchmark · llm agent secure-code-generation vulnerability-detection execution-grounded contamination-controlled

• CrackMeBench: Binary Reverse Engineering for Agents
  Ilia David and Arthur Gervais. arXiv 2026.
  Paper
  Evaluates tool-using agents on clean-room binary reverse-engineering challenges with deterministic grading.
  Emerging · Benchmark · llm agent malware-reverse-engineering decompilation binary dual-use execution-grounded contamination-controlled

• CREBench: Evaluating Large Language Models in Cryptographic Binary Reverse Engineering
  Baicheng Chen et al. COLM 2026.
  Paper · Code · Dataset · Project
  Evaluates LLMs across four levels of cryptographic binary reverse engineering, from algorithm identification to flag recovery.
  Emerging · Benchmark · llm binary cryptographic-reverse-engineering cryptography ctf-style execution-grounded

• CryptanalysisBench: Can LLMs do Cryptanalysis?
  Lukas Fluri et al. arXiv 2026.
  Paper · Code · Project
  Evaluates agents on 191 real cryptographic schemes across six primitive families using formal security games and fresh-randomness verification of submitted attack programs.
  Emerging · Benchmark · llm agent cryptanalysis cryptography dual-use execution-grounded adaptive-adversary

• CTIConnect: A Benchmark for Retrieval-Augmented LLMs over Heterogeneous Cyber Threat Intelligence
  Yutong Cheng et al. ACM SIGKDD 2026.
  Paper · Code · Dataset · Project
  Evaluates retrieval and reasoning across 1,859 expert-verified questions, nine tasks, and five heterogeneous cyber-threat-intelligence sources.
  Emerging · Benchmark · llm retrieval-augmented threat-intelligence attack-technique-mapping curated-real-data temporal

• CVE-Factory: Scaling Expert-Level Agentic Tasks for Code Security Vulnerability
  Xianzhen Luo et al. ICML 2026 Oral.
  Paper
  Builds executable environments from sparse CVE records and introduces LiveCVEBench with 190 tasks spanning 153 repositories and 14 programming languages.
  Emerging · Benchmark · llm agent multi-agent vulnerability-reproduction repository-scale execution-grounded

• CyberGym-E2E: Scalable Real-World Benchmark for AI Agents' End-to-End Cybersecurity Capabilities
  Tianneng Shi et al. ICML 2026.
  Paper
  Evaluates discovery, proof-of-concept generation, and patching across 920 vulnerabilities from 139 real open-source projects.
  Emerging · Benchmark · llm agent vulnerability-detection poc-generation patch-generation repository-scale execution-grounded

• CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale
  Zhun Wang et al. ICLR 2026.
  Paper · arXiv · Code · Dataset · Project
  Evaluates agents on generating executable PoCs that reproduce historical vulnerabilities in real, repository-scale software.
  Essential · Benchmark · llm agent source-code vulnerability-reproduction poc-generation dual-use real-world execution-grounded

• ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents
  Seunghyun Lee and David Brumley. arXiv 2026.
  Paper
  Grades sixteen progressive exploitation capabilities on 41 V8 bugs rather than equating a crash with code execution.
  Emerging · Benchmark · llm agent exploit-generation exploit-adaptation binary web offensive real-world

• ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?
  Zhun Wang et al. arXiv 2026.
  Paper · Code · Project
  Evaluates whether agents can extend vulnerability-triggering inputs into working exploits across userspace programs, V8, and the Linux kernel.
  Emerging · Benchmark · llm agent binary exploit-generation dual-use real-world execution-grounded

• Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks
  Songwen Zhao et al. ICML 2026.
  Paper · Code · Leaderboard
  Introduces SUSVIBES, 186 real feature-request tasks that jointly test the functionality and security of coding-agent changes to open-source repositories.
  Emerging · Benchmark · llm agent secure-code-generation vulnerability-detection repository-scale unit-tested

• REBENCH: A Procedural, Fair-by-Construction Benchmark for LLMs on Stripped-Binary Types and Names
  Jun Yeon Won et al. AIWare 2026.
  Paper
  Builds byte-aligned ground truth for fair evaluation of type and name recovery across architectures and optimization levels.
  Emerging · Benchmark · llm type-recovery function-naming binary dual-use real-world contamination-controlled

• REFORGE: A Method for Benchmarking LLMs' Reverse Engineering Capabilities in Decompiled Binary Function Naming
  Nicolas Koller and Andreas U. Schmidt. Applied Computing 2026.
  Paper
  Makes binary-to-source alignment uncertainty explicit when benchmarking LLM-based function naming.
  Emerging · Benchmark · llm function-naming decompilation binary dual-use synthetic contamination-controlled

• SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?
  Hwiwon Lee et al. arXiv 2026.
  Paper
  Benchmarks long-horizon PoC reproduction on 344 validated V8, SpiderMonkey, and Linux-kernel vulnerabilities.
  Emerging · Benchmark · llm agent vulnerability-reproduction poc-generation source-code repository dual-use real-world

• The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark
  Jeremy Spence et al. arXiv 2026.
  Paper
  Introduces 262 contamination-controlled binary instances from 19 private, real-world-scale programs with layered anti-analysis protections.
  Emerging · Benchmark · llm agent malware-reverse-engineering protocol-reverse-engineering decompilation binary firmware network

• Towards Effective Offensive Security LLM Agents: Hyperparameter Tuning, LLM as a Judge, and a Lightweight CTF Benchmark
  Minghao Shao et al. AAAI 2026.
  Paper · arXiv · Code
  Introduces CTFTiny, CTFJudge, and the CTF Competency Index for low-cost, fine-grained evaluation of offensive-security agents across fifty CTF challenges.
  Emerging · Benchmark · llm agent multi-agent ctf-solving attack-planning binary web cryptography offensive ctf-style model-graded

• Training Language Model Agents to Find Vulnerabilities with CTF-Dojo
  Terry Yue Zhuo et al. ICML 2026.
  Paper · arXiv
  Provides an interactive CTF environment and training recipe for agents that learn from executable security feedback.
  Emerging · Benchmark · llm agent reinforcement-learning ctf-solving vulnerability-detection binary web source-code

• BaxBench: Can LLMs Generate Correct and Secure Backends?
  Mark Vero et al. ICML 2025.
  Paper · Preprint · Code · Dataset · Leaderboard
  Evaluates secure backend generation across 392 tasks using comprehensive functional tests and executable end-to-end security exploits.
  Essential · Benchmark · llm secure-code-generation source-code web unit-tested execution-grounded

• Benchmarking LLMs and LLM-Based Agents in Practical Vulnerability Detection for Code Repositories
  Alperen Yildiz et al. ACL 2025.
  Paper
  Measures practical repository-level vulnerability detection beyond isolated-function classification.
  Recommended · Benchmark · llm agent vulnerability-detection vulnerability-localization source-code repository defensive real-world

• BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems
  Andy Zhang et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper · arXiv · Code
  Measures Detect, Exploit, and Patch capabilities on forty real bug bounties in twenty-five deployable systems and reports their historical dollar impact.
  Essential · Benchmark · llm agent vulnerability-detection exploit-generation patch-generation dual-use real-world execution-grounded

• Bridging Crypto with ML-based Solvers: the SAT Formulation and Benchmarks
  Xinhao Zheng et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper · Code
  Introduces SAT4CryptoBench for comparing standalone learned distinguishers, ML-guided solver heuristics, and hyperparameter optimization on cryptographic ANF and CNF instances.
  Recommended · Benchmark · classical-ml deep-learning neuro-symbolic cryptanalysis cryptography execution-grounded

• CVE-Bench: A Benchmark for AI Agents' Ability to Exploit Real-World Web Application Vulnerabilities
  Yuxuan Zhu et al. ICML 2025 Spotlight.
  Paper · arXiv · Code
  Provides reproducible sandboxes and automatic outcome checks for forty critical-severity web CVEs under zero-day and one-day information settings.
  Essential · Benchmark · llm agent web web-exploitation offensive real-world execution-grounded

• Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models
  Andy K. Zhang et al. ICLR 2025 Oral.
  Paper · arXiv · Code · Project
  Evaluates agents on forty professional CTF tasks from four recent competitions and adds expert-written subtasks for measuring partial progress on unsolved challenges.
  Essential · Benchmark · llm agent ctf-solving offensive ctf-style execution-grounded

• Decompile-Bench: Million-Scale Binary-Source Function Pairs for Real-World Binary Decompilation
  Hanzhuo Tan et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper · Preprint · Code
  Releases two million matched binary-source function pairs plus execution-oriented evaluation sets for training and comparing neural decompilers.
  Recommended · Benchmark · llm decompilation binary source-code repository-scale execution-grounded temporal

• DecompileBench: A Comprehensive Benchmark for Evaluating Decompilers in Real-World Scenarios
  Zeyu Gao et al. ACL Findings 2025.
  Paper
  Evaluates neural and traditional decompilers with semantic and execution-aware criteria on real programs.
  Recommended · Benchmark · llm decompilation binary source-code dual-use real-world execution-grounded

• JsDeObsBench: Measuring and Benchmarking LLMs for JavaScript Deobfuscation
  Guoqiang Chen, Xin Jin, and Zhiqiang Lin. ACM CCS 2025.
  Paper
  Introduces an execution-aware benchmark for measuring whether LLMs recover readable and behaviorally correct JavaScript from obfuscated programs.
  Recommended · Benchmark · llm decompilation source-code web execution-grounded

• SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks
  Hwiwon Lee et al. NeurIPS 2025.
  Paper
  Evaluates agents across real software-security workflows with automated, execution-grounded grading.
  Recommended · Benchmark · llm agent vulnerability-reproduction poc-generation patch-generation source-code repository dual-use

• SECODEPLT: A Unified Benchmark for Evaluating the Security Risks and Capabilities of Code GenAI
  Yuzhou Nie et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper · Code · Dataset
  Unifies insecure code generation, vulnerability detection, and patch generation across more than 5,900 examples, 44 CWE types, and four programming languages.
  Recommended · Benchmark · llm secure-code-generation vulnerability-detection patch-generation execution-grounded

• CTIBench: A Benchmark for Evaluating LLMs in Cyber Threat Intelligence
  Md Tanvirul Alam et al. NeurIPS 2024 Datasets and Benchmarks Spotlight.
  Paper · Code · Dataset
  Evaluates CTI knowledge, CWE mapping, CVSS prediction, ATT&CK extraction, and threat-actor attribution rather than generic cybersecurity question answering.
  Recommended · Benchmark · llm threat-intelligence attack-technique-mapping curated-real-data

• NYU CTF Bench: A Scalable Open-Source Benchmark Dataset for Evaluating LLMs in Offensive Security
  Minghao Shao et al. NeurIPS 2024 Datasets and Benchmarks.
  Paper · Code
  Provides 200 validated CSAW CTF challenges across six categories and an automated tool-using evaluation framework.
  Essential · Benchmark · llm agent ctf-solving offensive ctf-style execution-grounded

• VulDetectBench: Evaluating the Deep Capability of Vulnerability Detection with Large Language Models
  Yue Liu et al. arXiv 2024.
  Paper
  Separates shallow pattern matching from deeper vulnerability reasoning in LLM evaluation.
  Emerging · Benchmark · llm vulnerability-detection vulnerability-localization source-code defensive curated-real-data

• DiverseVul: A New Vulnerable Source Code Dataset for Deep Learning Based Vulnerability Detection
  Yizheng Chen et al. RAID 2023.
  Paper
  Builds a broad real-world vulnerable-function dataset and exposes generalization limits in learned detectors.
  Recommended · Dataset · deep-learning representation-learning vulnerability-detection source-code defensive real-world

• MOTIF: A Malware Reference Dataset with Ground Truth Family Labels
  Robert J. Joyce et al. Computers & Security 2022.
  Paper
  Provides expert-vetted malware-family labels to reduce noise in family-classification evaluation.
  Recommended · Dataset · classical-ml deep-learning family-attribution malware-classification binary defensive real-world

• SOREL-20M: A Large Scale Benchmark Dataset for Malicious PE Detection
  Richard Harang and Ethan M. Rudd. arXiv 2020.
  Paper · Code and Data
  Releases metadata, labels, behavioral tags, features, and baselines for twenty million Windows PE files, with disarmed malware binaries available under controlled terms.
  Recommended · Dataset · classical-ml deep-learning binary malware-detection real-world

• EMBER: An Open Dataset for Training Static PE Malware Machine Learning Models
  Hyrum S. Anderson and Phil Roth. arXiv 2018.
  Paper
  Releases a standard feature set, labels, and baseline for reproducible static Windows PE malware detection.
  Essential · Dataset · classical-ml malware-detection malware-classification binary defensive real-world

Papers by Security Task

This is the primary task-oriented view of the catalog: every paper has one canonical security-task placement even when it also appears in the benchmark section.

Vulnerability Discovery and Analysis

Papers in this section use AI to find, localize, classify, reproduce, prioritize, or dynamically uncover software vulnerabilities.

• BACAgent: LLM-Powered Detection of Broken-Access-Control Vulnerabilities in Web Applications
  Fengyu Liu et al. ACM CCS 2026.
  Paper
  Uses an LLM-powered agent to detect broken-access-control vulnerabilities in web applications.
  Emerging · System · llm agent vulnerability-detection web-exploitation web execution-grounded

• Cottontail: Large Language Model-Driven Concolic Execution for Highly Structured Test Input Generation
  Haoxin Tu et al. IEEE S&P 2026.
  Paper · Code
  Combines concolic execution with LLM constraint solving and history-guided seed acquisition to generate highly structured test inputs.
  Emerging · System · llm neuro-symbolic test-generation dynamic-analysis fuzzing execution-grounded

• CTX-Coder: Cross-Attention Architectures Empower LLMs for Long-Context Vulnerability Detection
  Jujie Wang et al. AAAI 2026.
  Paper
  Uses cross-attention over contextual functions and introduces CTX-VUL for repository-context vulnerability detection.
  Emerging · Method · llm vulnerability-detection vulnerability-localization source-code repository-scale

• CVE-Factory: Scaling Expert-Level Agentic Tasks for Code Security Vulnerability
  Xianzhen Luo et al. ICML 2026 Oral.
  Paper
  Builds executable vulnerability environments from CVE records and introduces LiveCVEBench across 153 repositories and 14 languages.
  Emerging · Benchmark · llm agent vulnerability-reproduction poc-generation repository-scale execution-grounded

• CyberGym-E2E: Scalable Real-World Benchmark for AI Agents' End-to-End Cybersecurity Capabilities
  Tianneng Shi et al. ICML 2026.
  Paper
  Evaluates discovery, proof-of-concept generation, and patching across 920 vulnerabilities from 139 real open-source projects.
  Emerging · Benchmark · llm agent vulnerability-detection vulnerability-reproduction patch-generation execution-grounded

• CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale
  Zhun Wang et al. ICLR 2026.
  Paper · arXiv · Code · Dataset · Project
  Evaluates agents on generating executable PoCs that reproduce historical vulnerabilities in real, repository-scale software.
  Essential · Benchmark · llm agent source-code vulnerability-reproduction poc-generation dual-use real-world execution-grounded

• deepSURF: Detecting Memory Safety Vulnerabilities in Rust Through Fuzzing LLM-Augmented Harnesses
  Georgios C. Androutsopoulos and Antonio Bianchi. IEEE S&P 2026.
  Paper · Code
  Combines static analysis and LLM-augmented fuzz harnesses for unsafe Rust, rediscovering 20 known bugs and finding six new vulnerabilities.
  Emerging · System · llm fuzzing test-generation source-code execution-grounded

• FirmAgent: Leveraging Fuzzing to Assist LLM Agents with IoT Firmware Vulnerability Discovery
  Jiangan Ji et al. NDSS 2026.
  Paper
  Grounds LLM-agent taint analysis and PoC refinement in fuzzing evidence, finding 182 vulnerabilities in 14 firmware images.
  Emerging · System · llm agent fuzzing firmware iot poc-generation execution-grounded

• From Documentation to Zero-day Vulnerabilities: LLM-Driven Fuzzing of JavaScript Engines in PDF Readers
  Suyue Guo et al. ACM CCS 2026.
  Paper · Code
  Uses LLM-derived grammars and API relationships plus constraint solving to generate complex JavaScript call sequences, finding 31 zero-day flaws in three PDF readers.
  Emerging · System · llm neuro-symbolic fuzzing test-generation source-code real-world execution-grounded

• MulVul: Retrieval-augmented Multi-Agent Code Vulnerability Detection via Cross-Model Prompt Evolution
  Zihan Wu et al. ACL 2026.
  Paper
  Combines CWE-specialized detector agents, vulnerability-knowledge retrieval, and cross-model prompt evolution across 130 weakness types.
  Emerging · System · llm multi-agent retrieval-augmented vulnerability-detection source-code

• PILOT: Command-line Interface Fuzzing via Path-Guided, Iterative Large Language Model Prompting
  Momoko Shiraishi, Yinzhi Cao, and Takahiro Shinagawa. IEEE S&P 2026.
  Paper
  Uses static call paths, iterative LLM prompting, and coverage feedback to generate meaningful option and file combinations for CLI fuzzing.
  Emerging · System · llm fuzzing test-generation vulnerability-detection real-world execution-grounded

• Revelio: Cost-Efficient Agentic Memory Safety Vulnerability Detection For Repository-Scale Codebases
  Yiwei Hou et al. arXiv 2026.
  Paper
  Combines inexpensive LLM-based triage, lightweight static analysis, and sanitizer-validated executable proofs to discover memory-safety vulnerabilities at repository scale.
  Emerging · System · llm agent vulnerability-detection vulnerability-reproduction poc-generation repository-scale real-world execution-grounded

• SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?
  Hwiwon Lee et al. arXiv 2026.
  Paper
  Benchmarks long-horizon PoC reproduction on 344 validated V8, SpiderMonkey, and Linux-kernel vulnerabilities.
  Emerging · Benchmark · llm agent vulnerability-reproduction poc-generation source-code repository dual-use real-world

• SpecAuditor: Generating Audit Specifications for LLM-Driven Bug Detection
  Miaoqian Lin and Hao Chen. IEEE S&P 2026.
  Paper · Code
  Extracts, generalizes, and retrieves project-specific audit specifications to guide LLM-based bug detection.
  Emerging · System · llm retrieval-augmented static-analysis vulnerability-detection repository-scale

• Specializing Language Models for Textual Fuzzing via Reinforcement Learning
  Jiayi Lin et al. IEEE S&P 2026.
  Paper
  Trains R1-Fuzz with coverage-derived questions and distance-based rewards, reporting higher coverage and 29 previously unknown vulnerabilities.
  Emerging · System · language-model reinforcement-learning fuzzing test-generation execution-grounded

• The Illusion of Rust Safety: Detecting Modular Unsafe Functions with LLMs
  Xiang Cheng et al. ACM CCS 2026.
  Paper
  Defines modular unsafe functions in Rust and uses fine-tuned LLMs to detect them and generate proof-of-concept witnesses.
  Emerging · System · llm vulnerability-detection poc-generation source-code repository-scale real-world

• VulAgent: Hypothesis-Validation Driven Multi-Agent Architecture for Vulnerability Detection
  Ziliang Wang et al. Findings of ACL 2026.
  Paper
  Structures auditing as explicit vulnerability hypotheses, trigger paths, and project-context validation to reduce speculative reports.
  Emerging · System · llm multi-agent vulnerability-detection vulnerability-localization repository

• Benchmarking LLMs and LLM-Based Agents in Practical Vulnerability Detection for Code Repositories
  Alperen Yildiz et al. ACL 2025.
  Paper
  Measures practical repository-level vulnerability detection beyond isolated-function classification.
  Recommended · Benchmark · llm agent vulnerability-detection vulnerability-localization source-code repository defensive real-world

• LLMxCPG: Context-Aware Vulnerability Detection Through Code Property Graph-Guided Large Language Models
  Ahmed Lekssays et al. USENIX Security 2025.
  Paper
  Uses code-property-graph slices to preserve vulnerability-relevant interprocedural context while reducing the amount of code presented to an LLM.
  Recommended · Method · llm graph-learning source-code vulnerability-detection static

• PromeFuzz: A Knowledge-Driven Approach to Fuzzing Harness Generation with Large Language Models
  Yuwei Liu et al. ACM CCS 2025.
  Paper
  Grounds LLM fuzz-harness generation in target-library knowledge and validates generated harnesses through execution.
  Recommended · System · llm retrieval-augmented fuzzing test-generation execution-grounded

• SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks
  Hwiwon Lee et al. NeurIPS 2025.
  Paper
  Evaluates agents across real software-security workflows with automated, execution-grounded grading.
  Recommended · Benchmark · llm agent vulnerability-reproduction poc-generation patch-generation source-code repository dual-use

• VulnLLM-R: Specialized Reasoning LLM with Agent Scaffold for Vulnerability Detection
  Yuzhou Nie et al. arXiv 2025.
  Paper · Code · Model
  Trains a compact security reasoning model and wraps it in an agent scaffold for project-level vulnerability detection across Python, C, C++, and Java.
  Emerging · Method · llm agent vulnerability-detection vulnerability-localization source-code repository-scale real-world

• YuraScanner: Leveraging LLMs for Task-driven Web App Scanning
  Aleksei Stafeev et al. NDSS 2025.
  Paper
  Uses task-driven LLM reasoning to steer dynamic web-application scanning toward security-relevant behavior.
  Recommended · System · llm agent dynamic-analysis vulnerability-detection web dual-use real-world execution-grounded

• Fuzz4All: Universal Fuzzing with Large Language Models
  Chunqiu Steven Xia et al. ICSE 2024.
  Paper
  Introduces a language-agnostic LLM fuzzing loop that generates and mutates inputs from natural-language specifications.
  Essential · System · llm fuzzing test-generation source-code dual-use execution-grounded

• Large Language Model Guided Protocol Fuzzing
  Ruijie Meng et al. NDSS 2024.
  Paper · Code
  Introduces ChatAFL, which extracts protocol grammars and state information from natural-language specifications and uses an LLM to enrich seeds and escape coverage plateaus.
  Essential · System · llm network fuzzing protocol-reverse-engineering execution-grounded real-world

• Large Language Models Are Edge-Case Generators: Crafting Unusual Programs for Fuzzing Deep Learning Libraries
  Yinlin Deng et al. ICSE 2024.
  Paper
  Uses historical bug-triggering programs and LLM prompting to generate unusual tests for deep-learning libraries.
  Recommended · System · llm fuzzing test-generation source-code dual-use execution-grounded

• ProphetFuzz: Fully Automated Prediction and Fuzzing of High-Risk Option Combinations with Only Documentation via Large Language Model
  Dawei Wang et al. ACM CCS 2024.
  Paper · Preprint · Code
  Predicts high-risk command-line option combinations from documentation and turns them into fully automated fuzzing campaigns.
  Recommended · System · llm fuzzing test-generation source-code real-world execution-grounded

• VulDetectBench: Evaluating the Deep Capability of Vulnerability Detection with Large Language Models
  Yue Liu et al. arXiv 2024.
  Paper
  Separates shallow pattern matching from deeper vulnerability reasoning in LLM evaluation.
  Emerging · Benchmark · llm vulnerability-detection vulnerability-localization source-code defensive curated-real-data

• DiverseVul: A New Vulnerable Source Code Dataset for Deep Learning Based Vulnerability Detection
  Yizheng Chen et al. RAID 2023.
  Paper
  Builds a broad real-world vulnerable-function dataset and exposes generalization limits in learned detectors.
  Recommended · Dataset · deep-learning representation-learning vulnerability-detection source-code defensive real-world

• Large Language Models Are Zero-Shot Fuzzers: Fuzzing Deep-Learning Libraries via Large Language Models
  Yinlin Deng et al. ISSTA 2023.
  Paper
  Shows that pretrained code models can synthesize and mutate valid tests for deep-learning libraries without task-specific training.
  Essential · System · llm fuzzing test-generation source-code dual-use execution-grounded

• LineVul: A Transformer-Based Line-Level Vulnerability Prediction
  Michael C. Fu and Chakkrit Tantithamthavorn. MSR 2022.
  Paper
  Applies Transformer representations and attention-based localization to function- and line-level vulnerability prediction.
  Recommended · Method · language-model representation-learning vulnerability-detection vulnerability-localization source-code defensive real-world static

• ReGVD: Revisiting Graph Neural Networks for Vulnerability Detection
  Van-Anh Nguyen et al. ICSE Companion 2022.
  Paper
  Revisits graph pooling and pretrained code representations for learned source-code vulnerability detection.
  Recommended · Method · graph-learning representation-learning vulnerability-detection source-code defensive static

• Montage: A Neural Network Language Model-Guided JavaScript Engine Fuzzer
  Suyoung Lee et al. USENIX Security 2020.
  Paper
  Learns fragment sequences from regression tests to generate semantically rich JavaScript-engine inputs.
  Essential · System · language-model deep-learning fuzzing test-generation source-code dual-use execution-grounded real-world

• Devign: Effective Vulnerability Identification by Learning Comprehensive Program Semantics via Graph Neural Networks
  Yaqin Zhou et al. NeurIPS 2019.
  Paper
  Combines syntax, control-flow, and data-flow relations in a graph neural network for function-level vulnerability classification on manually labeled real projects.
  Essential · Method · graph-learning representation-learning source-code vulnerability-detection static real-world

• NEUZZ: Efficient Fuzzing with Neural Program Smoothing
  Dongdong She et al. IEEE S&P 2019.
  Paper · Code
  Learns a smooth neural approximation of branching behavior so gradients can guide mutations through the otherwise discrete fuzzing search space.
  Essential · Method · deep-learning source-code binary fuzzing execution-grounded

• Compiler Fuzzing through Deep Learning
  Chris Cummins et al. ISSTA 2018.
  Paper
  Learns a generative model of code to produce compiler test cases that resemble human-written programs.
  Recommended · System · language-model deep-learning fuzzing test-generation source-code dual-use execution-grounded

• VulDeePecker: A Deep Learning-Based System for Vulnerability Detection
  Zhen Li et al. NDSS 2018.
  Paper · PDF
  Introduces code gadgets as a semantics-aware representation for learning vulnerability patterns and helped establish deep learning for source-code vulnerability detection.
  Essential · System · deep-learning source-code vulnerability-detection static

• Learn&Fuzz: Machine Learning for Input Fuzzing
  Patrice Godefroid, Hila Peleg, and Rishabh Singh. ASE 2017.
  Paper
  Uses a learned generative model of structured inputs to seed and guide coverage-oriented fuzzing.
  Essential · Method · language-model fuzzing test-generation source-code dual-use execution-grounded

Program Understanding and Reverse Engineering

Papers in this section recover security-relevant semantics from source code, bytecode, binaries, firmware, or undocumented protocols.

• CrackMeBench: Binary Reverse Engineering for Agents
  Ilia David and Arthur Gervais. arXiv 2026.
  Paper
  Evaluates tool-using agents on clean-room binary reverse-engineering challenges with deterministic grading.
  Emerging · Benchmark · llm agent malware-reverse-engineering decompilation binary dual-use execution-grounded contamination-controlled

• CREBench: Evaluating Large Language Models in Cryptographic Binary Reverse Engineering
  Baicheng Chen et al. COLM 2026.
  Paper · Code · Dataset · Project
  Evaluates LLMs across four levels of cryptographic binary reverse engineering, from algorithm identification to flag recovery.
  Emerging · Benchmark · llm binary cryptographic-reverse-engineering cryptography ctf-style execution-grounded

• Decompiling the Synergy: An Empirical Study of Human–LLM Teaming in Software Reverse Engineering
  Zion Leonahenahe Basque et al. NDSS 2026.
  Paper
  Combines a 153-practitioner survey with 109 hours of instrumented work from 48 participants to measure human–LLM reverse-engineering workflows.
  Emerging · Empirical Study · llm malware-reverse-engineering decompilation binary human-baseline

• Idioms: A Simple and Effective Framework for Turbo-Charging Local Neural Decompilation with Well-Defined Types
  Luke Dramko, Claire Le Goues, and Edward J. Schwartz. NDSS 2026.
  Paper
  Fine-tunes local LLMs to emit decompiled code with user-defined type definitions and introduces the REALTYPE dataset.
  Emerging · Method · llm decompilation type-recovery binary execution-grounded

• REBENCH: A Procedural, Fair-by-Construction Benchmark for LLMs on Stripped-Binary Types and Names
  Jun Yeon Won et al. AIWare 2026.
  Paper
  Builds byte-aligned ground truth for fair evaluation of type and name recovery across architectures and optimization levels.
  Emerging · Benchmark · llm type-recovery function-naming binary dual-use real-world contamination-controlled

• REFORGE: A Method for Benchmarking LLMs' Reverse Engineering Capabilities in Decompiled Binary Function Naming
  Nicolas Koller and Andreas U. Schmidt. Applied Computing 2026.
  Paper
  Makes binary-to-source alignment uncertainty explicit when benchmarking LLM-based function naming.
  Emerging · Benchmark · llm function-naming decompilation binary dual-use synthetic contamination-controlled

• Selective Knowledge Distillation: Fusing LLM Semantic Strengths with DNN Efficiency for Binary Code Similarity Detection
  Shize Zhou et al. ACL 2026.
  Paper
  Distills LLM-derived program semantics into efficient binary-similarity models using selective targets and discrepancy-weighted sampling.
  Emerging · Method · llm deep-learning representation-learning binary-similarity binary

• The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark
  Jeremy Spence et al. arXiv 2026.
  Paper
  Introduces 262 contamination-controlled binary instances from 19 private, real-world-scale programs with layered anti-analysis protections.
  Emerging · Benchmark · llm agent malware-reverse-engineering protocol-reverse-engineering decompilation binary firmware network

• Towards Generality: Task-Adaptive Binary Analysis via Semantic Retrieval and Verifiable Reasoning
  Yuzhe Liu et al. USENIX Security 2026.
  Paper
  Introduces BINREX, which grounds intent through semantic retrieval and executes verifiable IDAPython subtasks for general static-binary analysis.
  Emerging · System · llm agent retrieval-augmented malware-reverse-engineering binary execution-grounded

• Beyond Classification: Inferring Function Names in Stripped Binaries via Domain Adapted LLMs
  Linxi Jiang, Xin Jin, and Zhiqiang Lin. NDSS 2025.
  Paper
  Domain-adapts language models to infer meaningful names for functions in stripped binaries.
  Recommended · Method · llm function-naming decompilation binary dual-use real-world

• Decompile-Bench: Million-Scale Binary-Source Function Pairs for Real-World Binary Decompilation
  Hanzhuo Tan et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper · Preprint · Code
  Releases two million matched binary-source function pairs plus execution-oriented evaluation sets for training and comparing neural decompilers.
  Recommended · Benchmark · llm decompilation binary source-code repository-scale execution-grounded temporal

• DecompileBench: A Comprehensive Benchmark for Evaluating Decompilers in Real-World Scenarios
  Zeyu Gao et al. ACL Findings 2025.
  Paper
  Evaluates neural and traditional decompilers with semantic and execution-aware criteria on real programs.
  Recommended · Benchmark · llm decompilation binary source-code dual-use real-world execution-grounded

• Decompiling Smart Contracts with a Large Language Model
  Isaac David et al. arXiv 2025.
  Paper · Project
  Combines static EVM analysis with a fine-tuned language model to translate bytecode into readable Solidity with recovered names, control flow, and function signatures.
  Emerging · System · llm neuro-symbolic decompilation type-recovery function-naming bytecode smart-contract dual-use

• JsDeObsBench: Measuring and Benchmarking LLMs for JavaScript Deobfuscation
  Guoqiang Chen, Xin Jin, and Zhiqiang Lin. ACM CCS 2025.
  Paper
  Evaluates whether LLMs recover readable and behaviorally correct JavaScript from obfuscated programs with execution-aware criteria.
  Recommended · Benchmark · llm decompilation source-code web execution-grounded

• Nova: Generative Language Models for Assembly Code with Hierarchical Attention and Contrastive Learning
  Nan Jiang et al. ICLR 2025.
  Paper · Code
  Pretrains an assembly-language model with hierarchical attention and contrastive objectives over functions and compiler optimization levels.
  Recommended · Method · llm representation-learning binary-representation decompilation binary-similarity binary

• BinBert: Binary Code Understanding with a Fine-Tunable and Execution-Aware Transformer
  Fiorella Artuso et al. IEEE TDSC 2024.
  Paper
  Pretrains binary representations using execution-aware objectives and transfers them to multiple analysis tasks.
  Recommended · Method · language-model representation-learning binary-representation binary-similarity binary dual-use static

• LLM4Decompile: Decompiling Binary Code with Large Language Models
  Hanzhuo Tan et al. EMNLP 2024.
  Paper · Code
  Introduces an open model family trained to translate binary code into readable and executable source code.
  Essential · Method · llm decompilation code-representation binary source-code execution-grounded

• Augmenting Decompiler Output with Learned Variable Names and Types
  Qibin Chen et al. USENIX Security 2022 Distinguished Paper.
  Paper · Code
  Introduces DIRTY, a Transformer that restores meaningful variable names and types to decompiler output, together with a large paired source/decompilation dataset.
  Essential · System · language-model binary decompilation function-naming type-recovery

• jTrans: Jump-Aware Transformer for Binary Code Similarity
  Hao Wang et al. ISSTA 2022.
  Paper
  Injects control-flow jump structure into Transformer pretraining for large-scale binary similarity search.
  Recommended · Method · language-model representation-learning binary-similarity binary-representation binary dual-use real-world static

• PalmTree: Learning an Assembly Language Model for Instruction Embedding
  Xuezixiang Li, Yu Qu, and Heng Yin. ACM CCS 2021.
  Paper · Code
  Pretrains a contextual assembly-language model with instruction, control-flow, and data-dependency objectives for reuse across binary-analysis tasks.
  Recommended · Method · language-model representation-learning binary binary-representation

• TREX: Learning Execution Semantics from Micro-Traces for Binary Similarity
  Kexin Pei et al. IEEE TSE 2021.
  Paper
  Uses micro-traces as self-supervision so binary embeddings capture execution semantics across compilers and architectures.
  Recommended · Method · language-model representation-learning binary-similarity binary-representation binary dual-use execution-grounded

• XDA: Accurate, Robust Disassembly with Transfer Learning
  Kexin Pei et al. NDSS 2021.
  Paper
  Treats disassembly and function-boundary recovery as transfer learning over raw binary bytes.
  Essential · Method · language-model representation-learning function-identification binary-representation binary dual-use real-world static

• DeepBinDiff: Learning Program-Wide Code Representations for Binary Diffing
  Yue Duan et al. NDSS 2020.
  Paper
  Learns program-wide instruction embeddings and graph structure for cross-version binary diffing.
  Essential · System · graph-learning representation-learning binary-similarity software-provenance binary dual-use real-world static

• Asm2Vec: Boosting Static Representation Robustness for Binary Clone Search against Code Obfuscation and Compiler Optimization
  Steven H. H. Ding, Benjamin C. M. Fung, and Philippe Charland. IEEE S&P 2019.
  Paper
  Learns assembly-function embeddings designed to remain useful across optimization and obfuscation.
  Essential · Method · representation-learning binary-similarity binary-representation binary dual-use static real-world

• Coda: An End-to-End Neural Program Decompiler
  Cheng Fu et al. NeurIPS 2019.
  Paper
  Combines neural translation and error correction for end-to-end assembly-to-source decompilation.
  Essential · Method · deep-learning language-model decompilation binary source-code dual-use synthetic execution-grounded

• SAFE: Self-Attentive Function Embeddings for Binary Similarity
  Luca Massarelli et al. DIMVA 2019.
  Paper
  Builds efficient self-attentive instruction-sequence embeddings for cross-architecture function similarity.
  Recommended · Method · deep-learning representation-learning binary-similarity binary-representation binary dual-use static

• DEBIN: Predicting Debug Information in Stripped Binaries
  Jingxuan He et al. ACM CCS 2018.
  Paper
  Uses probabilistic structured prediction to recover variable names and types from stripped binaries.
  Essential · System · structured-prediction representation-learning function-naming type-recovery binary dual-use real-world static

• EKLAVYA: Recovering Function Types in Stripped Binaries
  Zheng Leong Chua et al. USENIX Security 2017.
  Paper
  Learns function signatures from disassembly to improve indirect-call resolution in stripped binaries.
  Essential · System · deep-learning type-recovery function-identification binary dual-use real-world static

• Neural Network-Based Graph Embedding for Cross-Platform Binary Code Similarity Detection
  Xiaojun Xu et al. ACM CCS 2017.
  Paper · PDF
  Introduces Gemini, a learned embedding of attributed control-flow graphs that made cross-architecture binary similarity substantially faster and more accurate than graph matching.
  Essential · Method · graph-learning representation-learning binary binary-similarity software-provenance

AI for Cryptography and Cryptanalysis

Papers in this section apply AI to cryptographic reasoning, cryptanalysis, solver guidance, side-channel analysis, and the recovery of cryptographic semantics from implementations.

• AICrypto: Evaluating Cryptography Capabilities of Large Language Models
  Yu Wang et al. ICML 2026.
  Paper · arXiv · Code · Dataset · Project
  Evaluates cryptographic knowledge, practical vulnerability exploitation, and proof reasoning with 135 multiple-choice questions, 150 CTF challenges, and 30 proof problems reviewed or constructed by cryptography experts.
  Emerging · Benchmark · llm agent cryptographic-reasoning cryptanalysis cryptography ctf-style human-baseline

• CREBench: Evaluating Large Language Models in Cryptographic Binary Reverse Engineering
  Baicheng Chen et al. COLM 2026.
  Paper · Code · Dataset · Project
  Evaluates LLMs across four levels of cryptographic binary reverse engineering, from algorithm identification to flag recovery.
  Emerging · Benchmark · llm binary cryptographic-reverse-engineering cryptography ctf-style execution-grounded

• CryptanalysisBench: Can LLMs do Cryptanalysis?
  Lukas Fluri et al. arXiv 2026.
  Paper · Code · Project
  Evaluates agents on 191 real cryptographic schemes across six primitive families using formal security games and fresh-randomness verification of submitted attack programs.
  Emerging · Benchmark · llm agent cryptanalysis cryptography dual-use execution-grounded adaptive-adversary

• Deep Neural Cryptography
  David Gerault et al. EUROCRYPT 2026.
  Paper · PDF · DOI
  Defines correctness and security for cryptographic primitives implemented as ReLU neural networks, breaks natural implementations using non-Boolean inputs, and gives a low-overhead provably secure construction.
  Recommended · Method · deep-learning cryptographic-implementation-analysis cryptographic-design-analysis cryptanalysis cryptography dual-use execution-grounded

• Halfspace Learning for Lattice Signature Key Recovery from Signs
  Marcus Brinkmann, Nicolai Kraus, and Alexander May. CRYPTO 2026.
  Paper · DOI · Code
  Models sign leakage as halfspace learning to recover HAWK, Falcon, and ML-DSA signing keys with strong noise tolerance.
  Emerging · Method · classical-ml side-channel-analysis cryptanalysis cryptography execution-grounded

• Neural-Inspired Advances in Integral Cryptanalysis
  Liu Zhang et al. EUROCRYPT 2026.
  Paper · DOI
  Uses neural distinguishers to discover integral features, interprets them with Boolean analysis, and converts them into improved classical attacks on SKINNY.
  Emerging · Method · deep-learning neural-cryptanalysis cryptanalysis cryptography execution-grounded

• AI for Code-based Cryptography
  Mohamed Malhou, Ludovic Perret, and Kristin Lauter. SAC 2025.
  Paper · Code
  Introduces a Transformer-based distinguisher for structured Goppa, MDPC, and QC-MDPC codes and studies whether learned representations expose structure relevant to post-quantum cryptanalysis.
  Recommended · Method · deep-learning language-model neural-cryptanalysis cryptographic-design-analysis cryptography dual-use

• Bridging Crypto with ML-based Solvers: the SAT Formulation and Benchmarks
  Xinhao Zheng et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper · Code
  Introduces SAT4CryptoBench for comparing standalone learned distinguishers, ML-guided solver heuristics, and hyperparameter optimization on cryptographic ANF and CNF instances.
  Recommended · Benchmark · classical-ml deep-learning neuro-symbolic cryptanalysis cryptography execution-grounded

• Is ML-Based Cryptanalysis Inherently Limited? Simulating Cryptographic Adversaries via Gradient-Based Methods
  Avital Shafran et al. CRYPTO 2024.
  Paper · DOI
  Formalizes sample- and gradient-based cryptanalytic adversaries and proves simulation results that clarify the expressive potential and limits of ML-based attacks.
  Recommended · Method · classical-ml deep-learning neural-cryptanalysis cryptanalysis cryptography

• SALSA VERDE: A Machine Learning Attack on Learning With Errors with Sparse Small Secrets
  Cathy Yuanchen Li et al. NeurIPS 2023.
  Paper
  Extends Transformer-based LWE attacks to larger dimensions and sparse binary, ternary, and narrow-Gaussian secrets.
  Essential · Method · deep-learning language-model neural-cryptanalysis cryptanalysis cryptography dual-use execution-grounded synthetic

• An Assessment of Differential-Neural Distinguishers
  Aron Gohr, Gregor Leander, and Patrick Neumann. IACR ePrint 2022.
  Paper
  Systematically reassesses what differential-neural distinguishers learn and corrects misleading multi-pair improvement claims.
  Recommended · Empirical Study · deep-learning neural-cryptanalysis cryptanalysis cryptography dual-use execution-grounded

• SALSA: Attacking Lattice Cryptography with Transformers
  Emily Wenger et al. NeurIPS 2022.
  Paper · Code
  Combines Transformer learning of modular arithmetic with statistical cryptanalysis to recover sparse binary secrets from small-to-medium LWE instances.
  Recommended · Method · deep-learning language-model neural-cryptanalysis cryptanalysis cryptography dual-use execution-grounded

• A Deeper Look at Machine Learning-Based Cryptanalysis
  Adrien Benamira et al. EUROCRYPT 2021.
  Paper
  Analyzes the internal behavior of neural distinguishers and relates learned signals to classical differential cryptanalysis.
  Essential · Empirical Study · deep-learning neural-cryptanalysis cryptanalysis cryptography dual-use execution-grounded

• SCAUL: Power Side-Channel Analysis with Unsupervised Learning
  Keyvan Ramezanpour, Paul Ampadu, and William Diehl. IEEE Transactions on Computers 2020.
  Paper
  Uses unsupervised representation learning to recover leakage features and keys from power traces without profiling labels.
  Recommended · Method · deep-learning unsupervised-learning side-channel-analysis cryptanalysis cryptography dual-use execution-grounded real-world

• Improving Attacks on Round-Reduced Speck32/64 Using Deep Learning
  Aron Gohr. CRYPTO 2019.
  Paper · PDF · Code
  Introduces deep residual neural distinguishers and a Bayesian key-search strategy that improve differential attacks on round-reduced Speck32/64.
  Essential · Method · deep-learning neural-cryptanalysis cryptanalysis cryptography dual-use execution-grounded

• Breaking Cryptographic Implementations Using Deep Learning Techniques
  Houssem Maghrebi, Thibault Portigliatti, and Emmanuel Prouff. SPACE 2016.
  Paper
  Demonstrates early deep-learning profiling attacks against masked cryptographic implementations.
  Essential · Method · deep-learning side-channel-analysis cryptanalysis cryptography dual-use execution-grounded real-world

Exploitation and Offensive Security

Papers in this section study AI systems that turn weaknesses into concrete impact or execute multi-step offensive workflows in controlled settings.

• AgentCyberRange: Benchmarking Frontier AI Systems in Realistic Cyber Ranges
  Fengyu Liu et al. arXiv 2026.
  Paper · Code · Web Dataset · Post-exploitation Dataset · Project
  Introduces an open multi-range benchmark with 110 vulnerabilities across 15 web applications and eight enterprise-like ranges containing 156 hosts, evaluating autonomous web exploitation and post-exploitation workflows.
  Emerging · Benchmark · llm agent automated-penetration-testing web-exploitation post-exploitation lateral-movement web network host offensive cyber-range execution-grounded

• Co-RedTeam: Orchestrated Security Discovery and Exploitation with LLM Agents
  Pengfei He et al. ICML 2026.
  Paper
  Orchestrates specialized agents for vulnerability discovery and exploitation with shared execution-grounded memory.
  Emerging · System · llm agent multi-agent automated-penetration-testing exploit-generation execution-grounded

• ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents
  Seunghyun Lee and David Brumley. arXiv 2026.
  Paper
  Grades sixteen progressive exploitation capabilities on 41 V8 bugs rather than equating a crash with code execution.
  Emerging · Benchmark · llm agent exploit-generation exploit-adaptation binary web offensive real-world

• ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?
  Zhun Wang et al. arXiv 2026.
  Paper · Code · Project
  Evaluates whether agents can extend vulnerability-triggering inputs into working exploits across userspace programs, V8, and the Linux kernel.
  Emerging · Benchmark · llm agent binary exploit-generation dual-use real-world execution-grounded

• Patch-to-PoC: A Systematic Study of Agentic LLM Systems for Linux Kernel N-Day Reproduction
  Jiahao Pu et al. arXiv 2026.
  Paper
  Studies whether agents can turn Linux-kernel patches into working n-day proofs of concept.
  Emerging · Empirical Study · llm agent exploit-adaptation poc-generation source-code repository dual-use real-world

• Teams of LLM Agents Can Exploit Zero-Day Vulnerabilities
  Yuxuan Zhu et al. EACL 2026.
  Paper · arXiv
  Studies coordinated multi-agent exploitation of previously unseen vulnerabilities in sandboxed systems.
  Emerging · Empirical Study · llm agent exploit-chaining automated-penetration-testing web offensive execution-grounded

• Towards Effective Offensive Security LLM Agents: Hyperparameter Tuning, LLM as a Judge, and a Lightweight CTF Benchmark
  Minghao Shao et al. AAAI 2026.
  Paper · arXiv · Code
  Introduces CTFTiny, CTFJudge, and the CTF Competency Index for low-cost, fine-grained evaluation of offensive-security agents across fifty CTF challenges.
  Emerging · Benchmark · llm agent multi-agent ctf-solving attack-planning binary web cryptography offensive ctf-style model-graded

• Training Language Model Agents to Find Vulnerabilities with CTF-Dojo
  Terry Yue Zhuo et al. ICML 2026.
  Paper · arXiv
  Provides an interactive CTF environment and training recipe for agents that learn from executable security feedback.
  Emerging · Benchmark · llm agent reinforcement-learning ctf-solving vulnerability-detection binary web source-code

• CVE-Bench: A Benchmark for AI Agents' Ability to Exploit Real-World Web Application Vulnerabilities
  Yuxuan Zhu et al. ICML 2025 Spotlight.
  Paper · Code
  Provides reproducible sandboxes and automatic outcome checks for forty critical-severity web CVEs under zero-day and one-day information settings.
  Essential · Benchmark · llm agent web web-exploitation offensive real-world execution-grounded

• Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models
  Andy K. Zhang et al. ICLR 2025 Oral.
  Paper · arXiv · Code · Project
  Evaluates agents on forty professional CTF tasks from four recent competitions and adds expert-written subtasks for measuring partial progress on unsolved challenges.
  Essential · Benchmark · llm agent ctf-solving offensive ctf-style execution-grounded

• Dynamic Risk Assessments for Offensive Cybersecurity Agents
  Boyi Wei et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper
  Shows how offensive-agent risk estimates change under adversarial improvement and additional inference-time compute on InterCode CTF.
  Recommended · Empirical Study · llm agent ctf-solving attack-planning adaptive-adversary execution-grounded

• EnIGMA: Interactive Tools Substantially Assist LM Agents in Finding Security Vulnerabilities
  Talor Abramovich et al. ICML 2025.
  Paper · arXiv · Project
  Shows that LM-friendly terminal interfaces, interactive debugger and connection tools, and output summarization materially improve agent performance across 390 CTF challenges.
  Essential · System · llm agent ctf-solving offensive ctf-style execution-grounded

• From Capabilities to Performance: Evaluating Key Functional Properties of LLM Architectures in Penetration Testing
  Lanxiao Huang et al. EMNLP 2025.
  Paper
  Isolates how memory, coordination, tool selection, planning, and monitoring affect singular and modular penetration-testing agents.
  Recommended · Empirical Study · llm agent multi-agent automated-penetration-testing attack-planning execution-grounded

• Measuring and Augmenting Large Language Models for Solving Capture-the-Flag Challenges
  Zimo Ji et al. ACM CCS 2025.
  Paper
  Measures LLM performance on capture-the-flag challenges and studies augmentations that improve tool-assisted solution workflows.
  Recommended · Empirical Study · llm agent ctf-solving attack-planning execution-grounded

• A Preliminary Study on Using Large Language Models in Software Pentesting
  Kumar Shashwat et al. WoS-CaD 2024.
  Paper
  Empirically probes how LLMs plan and execute multi-step software penetration-testing workflows.
  Emerging · Empirical Study · llm agent automated-penetration-testing attack-planning web host offensive execution-grounded

• LLM Agents Can Autonomously Exploit One-Day Vulnerabilities
  Richard Fang et al. arXiv 2024.
  Paper
  Measures autonomous exploitation when agents receive vulnerability descriptions for recently disclosed flaws.
  Emerging · Empirical Study · llm agent exploit-adaptation automated-penetration-testing web offensive execution-grounded synthetic

• LLM Agents Can Autonomously Hack Websites
  Richard Fang et al. arXiv 2024.
  Paper
  Provides an early execution-grounded evaluation of autonomous LLM agents attacking vulnerable web applications.
  Emerging · Empirical Study · llm agent web-exploitation automated-penetration-testing web offensive execution-grounded synthetic

• NYU CTF Bench: A Scalable Open-Source Benchmark Dataset for Evaluating LLMs in Offensive Security
  Minghao Shao et al. NeurIPS 2024 Datasets and Benchmarks.
  Paper · Code
  Provides 200 validated CSAW CTF challenges across six categories and an automated tool-using evaluation framework.
  Essential · Benchmark · llm agent ctf-solving offensive ctf-style execution-grounded

• PentestGPT: Evaluating and Harnessing Large Language Models for Automated Penetration Testing
  Gelei Deng et al. USENIX Security 2024, Distinguished Artifact Award.
  Paper · Code
  Introduces a penetration-testing task tree and a modular LLM system that separates reasoning, command generation, and output parsing to reduce context loss during long workflows.
  Essential · System · llm agent automated-penetration-testing offensive ctf-style execution-grounded

Remediation and Secure Software Engineering

Papers in this section use AI to generate, validate, or prioritize security patches and to improve secure software development.

• AutoBaxBuilder: Bootstrapping Code Security Benchmarking
  Tobias von Arx et al. ICML 2026.
  Paper
  Automates construction of code-security tasks with functional tests and executable exploits while reducing benchmark contamination.
  Emerging · Benchmark · llm agent secure-code-generation vulnerability-detection execution-grounded

• GoodVibe: Security-by-Vibe for LLM-Based Code Generation
  Maximilian Thang et al. USENIX Security 2026.
  Paper
  Uses neuron attribution and selective fine-tuning to improve secure code generation across C++, Java, Swift, and Go.
  Emerging · Method · llm secure-code-generation security-code-review source-code unit-tested

• Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks
  Songwen Zhao et al. ICML 2026.
  Paper · Code · Leaderboard
  Jointly evaluates functionality and security on 186 real feature-request tasks with vulnerable human implementations.
  Emerging · Benchmark · llm agent secure-code-generation vulnerability-detection repository-scale

• Kintsugi: Empowering LLMs to Mitigate Web Vulnerabilities via Runtime Policy Injection
  Yihao Peng et al. USENIX Security 2026.
  Paper
  Uses LLM-localized vulnerable code and differential syscall analysis to inject temporary least-privilege containment policies.
  Emerging · System · llm patch-generation patch-validation security-specification web execution-grounded

• PORTGPT: Towards Automated Backporting Using Large Language Models
  Zhaoyang Li et al. IEEE S&P 2026.
  Paper · Code
  Builds a tool-using agent that backports security patches using repository history and compiler feedback, including nine patches merged into Linux.
  Emerging · System · llm agent patch-generation patch-validation repository real-world execution-grounded

• SecCodePRM: A Process Reward Model for Code Security
  Weichen Yu et al. ICML 2026.
  Paper
  Introduces step-level process rewards for identifying insecure reasoning trajectories and improving secure code generation.
  Emerging · Method · llm reinforcement-learning secure-code-generation security-code-review source-code

• SecPI: Secure Code Generation with Reasoning Models via Security Reasoning Internalization
  Hao Wang et al. arXiv 2026.
  Paper · Code
  Fine-tunes reasoning models to internalize structured security analysis and generate safer code without requiring security instructions at inference time.
  Emerging · Method · llm secure-code-generation security-code-review source-code unit-tested execution-grounded

• APPATCH: Automated Adaptive Prompting Large Language Models for Real-World Software Vulnerability Patching
  Yu Nong et al. USENIX Security 2025.
  Paper
  Combines semantics-aware scoping, adaptive exemplar selection, and cross-model validation to patch real vulnerabilities without requiring an exploit input or model fine-tuning.
  Recommended · System · llm source-code patch-generation real-world

• BaxBench: Can LLMs Generate Correct and Secure Backends?
  Mark Vero et al. ICML 2025.
  Paper · Preprint · Code · Dataset · Leaderboard
  Evaluates secure backend generation across 392 tasks using comprehensive functional tests and executable end-to-end security exploits.
  Essential · Benchmark · llm secure-code-generation source-code web unit-tested execution-grounded

• BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems
  Andy Zhang et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper · Code
  Measures Detect, Exploit, and Patch capabilities on forty real bug bounties in twenty-five deployable systems and reports their historical dollar impact.
  Essential · Benchmark · llm agent vulnerability-detection exploit-generation patch-generation dual-use real-world execution-grounded

• Give LLMs a Security Course: Securing Retrieval-Augmented Code Generation via Knowledge Injection
  Bo Lin et al. ACM CCS 2025.
  Paper
  Injects security knowledge into retrieval-augmented code generation to reduce vulnerable outputs while retaining functional quality.
  Recommended · Method · llm retrieval-augmented secure-code-generation security-code-review source-code

• PATCHAGENT: A Practical Program Repair Agent Mimicking Human Expertise
  Zheng Yu et al. USENIX Security 2025.
  Paper
  Integrates fault localization, patch generation, building, PoC validation, and regression testing in a tool-using repair agent evaluated on 178 real-world vulnerabilities.
  Essential · System · llm agent source-code patch-generation patch-validation real-world execution-grounded

• PurpCode: Reasoning for Safer Code Generation
  Jiawei Liu et al. NeurIPS 2025.
  Paper
  Trains code models to reason over security rules and produce safer implementations while preserving general coding utility.
  Recommended · Method · llm reinforcement-learning secure-code-generation security-code-review unit-tested

• SECODEPLT: A Unified Benchmark for Evaluating the Security Risks and Capabilities of Code GenAI
  Yuzhou Nie et al. NeurIPS 2025 Datasets and Benchmarks.
  Paper · Code · Dataset
  Unifies insecure code generation, vulnerability detection, and patch generation with dynamic security artifacts across four languages.
  Recommended · Benchmark · llm secure-code-generation vulnerability-detection patch-generation execution-grounded

• A Case Study of LLM for Automated Vulnerability Repair: Assessing Impact of Reasoning and Patch Validation Feedback
  Ummay Kulsum et al. AIware 2024.
  Paper
  Combines LLM patch generation with program analysis and iterative validation for real vulnerabilities.
  Recommended · System · llm agent patch-generation patch-validation program-repair source-code defensive execution-grounded

• SecCoder: Towards Generalizable and Robust Secure Code Generation
  Boyu Zhang et al. EMNLP 2024.
  Paper
  Trains code models to generate secure implementations that generalize across weakness patterns and prompts.
  Recommended · Method · llm secure-code-generation security-code-review source-code defensive static

• InferFix: End-to-End Program Repair with LLMs
  Matthew Jin et al. ESEC/FSE 2023.
  Paper
  Combines retrieval and a repair model to generate fixes for real-world defects and security vulnerabilities.
  Recommended · System · llm program-repair patch-generation source-code defensive real-world execution-grounded

• Neural Transfer Learning for Repairing Security Vulnerabilities in C Code
  Zimin Chen, Steve Kommrusch, and Martin Monperrus. IEEE Transactions on Software Engineering 2023.
  Paper · arXiv · Code
  Introduces VRepair, which transfers knowledge from a large bug-fix corpus to the smaller-data problem of generating security patches for C vulnerabilities.
  Recommended · Method · deep-learning language-model source-code patch-generation program-repair

• VulRepair: A T5-Based Automated Software Vulnerability Repair
  Michael C. Fu et al. ESEC/FSE 2022.
  Paper
  Fine-tunes T5 for vulnerability repair and evaluates exact and semantically equivalent fixes on real patches.
  Recommended · Method · language-model patch-generation program-repair source-code defensive real-world execution-grounded

Malware and Unwanted Software

Papers in this section use AI to detect, classify, attribute, understand, or explain malware and other unwanted software.

• Beyond the TESSERACT: Trustworthy Dataset Curation for Sound Evaluations of Android Malware Classifiers
  Theo Chow et al. IEEE SaTML 2026.
  Paper
  Identifies five overlooked dataset-curation factors that materially change evaluations of learning-based Android malware classifiers.
  Emerging · Empirical Study · classical-ml deep-learning malware-detection malware-classification mobile temporal

• MalTree: Tracing Malware Evolution using Embeddings at Scale
  Akash Amalan, Georgios Smaragdakis, and Tom Viering. ICML 2026.
  Paper
  Builds phylogeny-inspired malware evolution trees from structural, behavioral, and image embeddings at a scale exceeding 100,000 samples.
  Emerging · Method · deep-learning representation-learning family-attribution behavior-analysis binary temporal

• Threat2Traffic: Multi-Agent Environment Synthesis for Malware Traffic Generation from Threat Intelligence
  Haoyang Chen et al. ICML 2026.
  Paper · Code
  Synthesizes malware execution environments from threat reports to reproduce network traffic for eight malware families.
  Emerging · System · llm multi-agent sandbox-analysis traffic-classification threat-intelligence execution-grounded

• Beyond Raw Bytes: Towards Large Malware Language Models
  Luke Kurlandski et al. NDSS 2025.
  Paper
  Pretrains large language-style models over semantically structured malware representations rather than raw bytes alone.
  Recommended · Method · language-model representation-learning malware-detection malware-classification binary defensive real-world

• FlowMalTrans: Unsupervised Binary Code Translation for Malware Detection Using Flow-Adapter Architecture
  Minghao Hu et al. Findings of EMNLP 2025.
  Paper
  Translates binary representations across instruction-set architectures so a malware detector trained on one architecture can transfer to another.
  Recommended · Method · deep-learning unsupervised-learning malware-detection binary iot

• MOTIF: A Malware Reference Dataset with Ground Truth Family Labels
  Robert J. Joyce et al. Computers & Security 2022.
  Paper
  Provides expert-vetted malware-family labels to reduce noise in family-classification evaluation.
  Recommended · Dataset · classical-ml deep-learning family-attribution malware-classification binary defensive real-world

• Neurlux: Dynamic Malware Analysis without Feature Engineering
  Chani Jindal et al. ACSAC 2021.
  Paper
  Learns directly from tokenized sandbox reports for dynamic malware detection without hand-engineered features.
  Recommended · Method · language-model deep-learning behavior-analysis malware-detection binary security-telemetry defensive real-world

• SOREL-20M: A Large Scale Benchmark Dataset for Malicious PE Detection
  Richard Harang and Ethan M. Rudd. arXiv 2020.
  Paper · Code and Data
  Releases metadata, labels, behavioral tags, features, and baselines for twenty million Windows PE files, with disarmed malware binaries available under controlled terms.
  Recommended · Dataset · classical-ml deep-learning binary malware-detection real-world

• EMBER: An Open Dataset for Training Static PE Malware Machine Learning Models
  Hyrum S. Anderson and Phil Roth. arXiv 2018.
  Paper
  Releases a standard feature set, labels, and baseline for reproducible static Windows PE malware detection.
  Essential · Dataset · classical-ml malware-detection malware-classification binary defensive real-world

• Malware Detection by Eating a Whole EXE
  Edward Raff et al. AAAI 2018 Workshop on Artificial Intelligence for Cyber Security.
  Paper
  Introduces MalConv and frames end-to-end learning directly from multi-million-byte Windows executables as a practical sequence-modeling problem.
  Essential · Method · deep-learning binary malware-detection static

• MaMaDroid: Detecting Android Malware by Building Markov Chains of Behavioral Models
  Enrico Mariconti et al. NDSS 2017.
  Paper
  Models Android API-call abstractions as Markov chains for behavioral malware detection resilient to API evolution.
  Essential · System · classical-ml mobile-malware malware-detection behavior-analysis mobile bytecode defensive real-world

• Transcend: Detecting Concept Drift in Malware Classification Models
  Roberto Jordaney et al. USENIX Security 2017.
  Paper
  Uses conformal evaluation to identify when evolving malware distributions make a classifier unreliable.
  Essential · Method · classical-ml malware-detection binary defensive real-world temporal

• DREBIN: Effective and Explainable Detection of Android Malware in Your Pocket
  Daniel Arp et al. NDSS 2014.
  Paper
  Introduces lightweight and explainable on-device Android malware detection over broad static features.
  Essential · System · classical-ml mobile-malware malware-detection mobile bytecode defensive real-world static

Network, Host, and Infrastructure Defense

Papers in this section apply AI to defensive analysis of network traffic, system logs, host activity, provenance, cloud systems, and other operational telemetry.

• ConCap: Practical Network Traffic Generation for (ML- and) Flow-based Intrusion Detection Systems
  Miel Verkerken et al. IEEE SaTML 2026.
  Paper
  Generates automatically labeled packets and flows in a configurable isolated network for reproducible intrusion-detection experiments.
  Emerging · System · classical-ml deep-learning intrusion-detection traffic-classification network real-world

• Cloak, Honey, Trap: Proactive Defenses Against LLM Agents
  Daniel Ayzenshteyn, Roy Weiss, and Yisroel Mirsky. USENIX Security 2025.
  Paper · PDF · Code
  Introduces six strategies and fifteen techniques for detecting, delaying, and stopping LLM-powered attack agents through cloaking, honeytokens, traps, and deception, with an open-source implementation evaluated on CTF machines.
  Essential · System · llm agent defensive autonomous-ai-agent cyber-deception honeypot-and-honeytoken ctf-style execution-grounded

• Self-Supervised Learning of Graph Representations for Network Intrusion Detection
  Lorenzo Guerra et al. NeurIPS 2025.
  Paper
  Uses masked self-supervised graph representation learning to model flow context for network intrusion detection.
  Recommended · Method · graph-learning representation-learning unsupervised-learning intrusion-detection network

• DeepCASE: Semi-Supervised Contextual Analysis of Security Events
  Thijs van Ede et al. IEEE S&P 2022.
  Paper
  Learns contextual event sequences and analyst feedback to prioritize large volumes of security events.
  Essential · System · deep-learning semi-supervised-learning log-analysis alert-triage host security-telemetry defensive real-world

• ET-BERT: A Contextualized Datagram Representation with Pre-Training Transformers for Encrypted Traffic Classification
  Xinjie Lin et al. The Web Conference 2022.
  Paper
  Pretrains contextual datagram representations for encrypted-traffic classification with limited labels.
  Essential · Method · language-model representation-learning traffic-classification intrusion-detection network defensive real-world

• ShadeWatcher: Recommendation-Guided Cyber Threat Analysis Using System Audit Records
  Jun Zeng et al. IEEE S&P 2022.
  Paper
  Frames provenance-based threat analysis as recommendation over audit-record entities and behaviors.
  Essential · System · graph-learning representation-learning host-based-detection threat-hunting host security-telemetry defensive real-world

• UNICORN: Runtime Provenance-Based Detector for Advanced Persistent Threats
  Xueyuan Han et al. NDSS 2020.
  Paper
  Builds streaming graph sketches of system provenance to detect long-running advanced persistent threats.
  Essential · System · graph-learning streaming-learning host-based-detection threat-hunting host security-telemetry defensive real-world

• Kitsune: An Ensemble of Autoencoders for Online Network Intrusion Detection
  Yisroel Mirsky et al. NDSS 2018.
  Paper · PDF · Code
  Uses an ensemble of small autoencoders for unsupervised, online network anomaly detection that can train and run on resource-constrained gateways.
  Essential · System · deep-learning network intrusion-detection network-anomaly-detection real-world

• DeepLog: Anomaly Detection and Diagnosis from System Logs through Deep Learning
  Min Du and Feifei Li. ACM CCS 2017.
  Paper
  Models log-key sequences with recurrent networks for online anomaly detection and workflow diagnosis.
  Essential · Method · deep-learning language-model log-analysis network-anomaly-detection host security-telemetry defensive real-world

• BotMiner: Clustering Analysis of Network Traffic for Protocol- and Structure-Independent Botnet Detection
  Guofei Gu et al. USENIX Security 2008.
  Paper
  Correlates communication and malicious-activity clusters to detect botnets independently of protocol structure.
  Essential · System · unsupervised-learning classical-ml botnet-detection traffic-classification network defensive real-world

• BotSniffer: Detecting Botnet Command and Control Channels in Network Traffic
  Guofei Gu, Junjie Zhang, and Wenke Lee. NDSS 2008.
  Paper
  Detects botnet command-and-control channels through spatial-temporal correlation in network behavior.
  Essential · System · classical-ml botnet-detection traffic-classification network defensive real-world

Social Engineering, Identity, and Abuse

Papers in this section analyze or detect phishing, malicious services, credential abuse, impersonation, fraud, and other attacks on people or online trust.

• When LLMs Go Online: The Emerging Threat of Web-Enabled LLMs
  Hanna Kim et al. USENIX Security 2025.
  Paper
  Studies malicious web-enabled LLM services and the risks created when models can browse and act online.
  Recommended · Empirical Study · llm agent online-abuse social-engineering web offensive real-world

• KnowPhish: Large Language Models Meet Multimodal Knowledge Graphs for Enhancing Reference-Based Phishing Detection
  Yuexin Li et al. USENIX Security 2024.
  Paper
  Builds a multimodal knowledge base covering twenty thousand brands and uses an LLM to extract textual brand evidence, extending reference-based detectors beyond logo-only pages.
  Recommended · System · llm multimodal-model security-knowledge-graph web phishing-detection real-world

• Malla: Demystifying Real-World Large Language Model Integrated Malicious Services
  Zilong Lin et al. USENIX Security 2024.
  Paper
  Measures how underground services operationalize LLMs for phishing, malware, and other malicious content.
  Essential · Empirical Study · llm online-abuse phishing-generation social-engineering web email offensive real-world

• Inferring Phishing Intention via Webpage Appearance and Dynamics: A Deep Vision Based Approach
  Ruofan Liu et al. USENIX Security 2022.
  Paper
  Combines visual brand recognition with active webpage interaction to infer credential-stealing intent.
  Essential · System · computer-vision deep-learning phishing-detection web defensive real-world execution-grounded

• Phishpedia: A Hybrid Deep Learning Based Approach to Visually Identify Phishing Webpages
  Yun Lin et al. USENIX Security 2021.
  Paper · Code
  Combines logo detection with Siamese similarity matching to identify the impersonated brand and provide a visual explanation without training on phishing samples.
  Essential · System · deep-learning multimodal-model web phishing-detection real-world

• URLNet: Learning a URL Representation with Deep Learning for Malicious URL Detection
  Hung Le et al. arXiv 2018.
  Paper
  Learns character- and word-level URL representations for scalable malicious-link detection.
  Essential · Method · deep-learning representation-learning malicious-url-detection phishing-detection web defensive real-world

Threat Intelligence, SOC, and Incident Response

Papers in this section help analysts extract threat intelligence, map adversary behavior, triage security events, hunt threats, investigate incidents, or coordinate response.

• CTIConnect: A Benchmark for Retrieval-Augmented LLMs over Heterogeneous Cyber Threat Intelligence
  Yutong Cheng et al. ACM SIGKDD 2026.
  Paper · Code · Dataset · Project
  Evaluates retrieval and reasoning across 1,859 expert-verified questions, nine tasks, and five heterogeneous cyber-threat-intelligence sources.
  Emerging · Benchmark · llm retrieval-augmented threat-intelligence attack-technique-mapping curated-real-data temporal

• LLMs in the SOC: An Empirical Study of Human-AI Collaboration in Security Operations Centres
  Ronal Singh et al. IEEE S&P 2026.
  Paper
  Analyzes 3,090 queries from 45 analysts over ten months to characterize how LLMs are used in an operational security operations center.
  Emerging · Empirical Study · llm soc-assistance incident-response human-ai-collaboration security-telemetry defensive real-world

• RedSage: A Cybersecurity Generalist LLM
  Naufal Suryanto et al. ICLR 2026.
  Paper · Preprint
  Develops and evaluates a cybersecurity-specialized generalist language model across knowledge, reasoning, threat intelligence, and applied security tasks.
  Emerging · Model · llm security-foundation-model threat-intelligence security-question-answering defensive curated-real-data

• Toward Cybersecurity-Expert Small Language Models
  Matan Levi et al. ICML 2026.
  Paper
  Introduces CyberPal 2.0, a family of compact cybersecurity-specialized language models trained and evaluated for practical expert-level security assistance.
  Emerging · Model · llm small-language-model security-foundation-model security-question-answering defensive curated-real-data

• TxRay: Agentic Postmortem of Live Blockchain Attacks
  Ziyue Wang et al. SBC 2026.
  Paper · Project
  Reconstructs real DeFi attack lifecycles from seed transactions, derives evidence-backed root causes, and generates self-validating executable proofs of concept.
  Emerging · System · llm agent multi-agent incident-investigation root-cause-analysis poc-generation blockchain dual-use execution-grounded

• CTINexus: Automatic Cyber Threat Intelligence Knowledge Graph Construction Using Large Language Models
  Yutong Cheng et al. IEEE EuroS&P 2025.
  Paper · arXiv · Code · Project
  Uses optimized in-context learning, hierarchical entity alignment, and relation prediction to construct cybersecurity knowledge graphs from unstructured threat reports.
  Recommended · System · llm nlp threat-intelligence-extraction entity-relation-extraction security-knowledge-graph curated-real-data

• OCR-APT: Reconstructing APT Stories from Audit Logs using Subgraph Anomaly Detection and LLMs
  Ahmed Aly, Essam Mansour, and Amr Youssef. ACM CCS 2025.
  Paper
  Combines anomalous provenance subgraphs with LLM reasoning to reconstruct coherent advanced-persistent-threat narratives from audit logs.
  Recommended · System · llm graph-learning incident-investigation threat-hunting host security-telemetry defensive real-world

• SoK: Automated TTP Extraction from CTI Reports – Are We There Yet?
  Marvin Büchel et al. USENIX Security 2025.
  Paper
  Unifies and reevaluates more than forty NLP-based approaches to extracting ATT&CK tactics, techniques, and procedures from threat reports, exposing dataset, ontology, and comparability problems.
  Essential · SoK · language-model llm threat-intelligence attack-technique-mapping real-world

• CTIBench: A Benchmark for Evaluating LLMs in Cyber Threat Intelligence
  Md Tanvirul Alam et al. NeurIPS 2024 Datasets and Benchmarks Spotlight.
  Paper · Code · Dataset
  Evaluates CTI knowledge, CWE mapping, CVSS prediction, ATT&CK extraction, and threat-actor attribution rather than generic cybersecurity question answering.
  Recommended · Benchmark · llm threat-intelligence attack-technique-mapping curated-real-data

• ThreatKG: An AI-Powered System for Automated Open-Source Cyber Threat Intelligence Gathering and Management
  Peng Gao et al. LAMPS@ACM CCS 2024.
  Paper · arXiv
  Automates collection of public threat reports, AI-based entity and relation extraction, and continuous integration into an extensible threat knowledge graph.
  Recommended · System · deep-learning nlp structured-prediction threat-intelligence-extraction entity-relation-extraction security-knowledge-graph real-world

• EXTRACTOR: Extracting Attack Behavior from Threat Reports
  Kiavash Satvat, Rigel Gjomemo, and V. N. Venkatakrishnan. EuroS&P 2021.
  Paper
  Extracts attack behaviors and causal provenance graphs from unstructured threat reports for threat hunting.
  Essential · System · language-model threat-intelligence-extraction entity-relation-extraction threat-hunting threat-intelligence security-telemetry defensive real-world

• TTPDrill: Automatic and Accurate Extraction of Threat Actions from Unstructured Text of CTI Sources
  Ghaith Husari et al. ACSAC 2017.
  Paper
  Introduces ontology-guided NLP for extracting threat actions and mapping them to tactics, techniques, and kill-chain phases.
  Essential · System · nlp classical-ml threat-intelligence-extraction attack-technique-mapping threat-intelligence defensive real-world

Defending Against AI-Enabled Cyber Attacks

This section covers defenses for conventional software, networks, and infrastructure when AI systems or autonomous agents act as attackers or attack tools. It does not cover protecting AI models or agents themselves, which remains outside the scope of this list.

• Cloak, Honey, Trap: Proactive Defenses Against LLM Agents
  Daniel Ayzenshteyn, Roy Weiss, and Yisroel Mirsky. USENIX Security 2025.
  Paper · PDF · Code
  Introduces six strategies and fifteen techniques for detecting, delaying, and stopping LLM-powered attack agents through cloaking, honeytokens, traps, and deception, with an open-source implementation evaluated on CTF machines.
  Essential · System · llm agent defensive autonomous-ai-agent cyber-deception honeypot-and-honeytoken ctf-style execution-grounded

Curation Statement

This section explains the evidence and editorial judgment used to decide which papers belong in the list.

Venue prestige, citation count, GitHub popularity, and social-media attention are useful signals, but none is sufficient on its own. Every entry must have a clear cybersecurity contribution and a documented reason to read it. Community attention is recorded only when it can be linked to an author thread, independent technical discussion, or another verifiable source.

Contributing

Contributions are highly welcome, especially missing foundational papers, important recent work, corrected publication metadata, and stronger descriptions of a paper's contribution or limitations.

Before submitting, please read the scope, inclusion criteria, and contribution guidelines. You can propose a paper through the paper-suggestion issue form or open a pull request that updates both the canonical YAML record and its README presentation.

License and Citation

This repository is available under the MIT License. If this catalog supports your research, please cite the repository using the metadata in CITATION.cff.