Evaluating Open-Weight LLMs for Generating Structured Threat Information for Autonomous Vehicle Vulnerabilities
arXiv:2607.16175v1 [cs.CR] 17 Jul 2026
Md Erfan, Ahmed Ryan, Md Kamal Hossain Chowdhury and Md Rayhanur Rahman† Abstract—Connected and Autonomous Vehicles (CAVs) rely on interconnected software and hardware components, including sensors, Electronic Control Units, in-vehicle infotainment systems, and telematics units, where vulnerabilities can compromise assets, users, and vehicle operations. These vulnerabilities are commonly documented as plain text in the Common Vulnerabilities and Exposures (CVE) database; however, security practitioners require structured information about affected assets, types of weaknesses, and attack behaviors to effectively mitigate the risks from these vulnerabilities. To this end, we evaluate openweight Large Language Models (LLMs) for generating Structured Threat Information Expression (STIX), a well-known structured format for representing threat information, for CAV-related CVEs. We construct a dataset called CAV-STIXGen that maps CAV vulnerability descriptions to STIX domain objects (SDO), STIX relationship objects (SRO), Common Weakness Enumeration (CWE), and MITRE ATT&CK techniques mappings. Using this dataset, we evaluated 11 open-weight LLMs (4B to 120B parameters) across various prompting strategies and temperatures. Single-model configurations achieve F1 scores of 0.94 for SDO, 0.63 for SRO, and 0.99 for CWE mapping, while complete MITRE ATT&CK mapping remains challenging. In a multi-agent setup, Gemma-4-31B and Codestral-22B achieve F1 scores of 0.91 for SDOs and 0.43 for SROs, respectively. Lastly, we analyze CWE and MITRE ATT&CK co-occurrences to identify recurring threat patterns in the CAV domain, demonstrating how AI-assisted vulnerability-to-STIX translation can automate threat intelligence and prioritize defense in transportation security. Index Terms—STIX Domain Objects, STIX Relationship Objects, CWE, Mitre Attack Techniques
I. I NTRODUCTION Connected and Autonomous Vehicles (CAV) operate through a complex interplay of software and hardware modules, such as sensors (e.g., LiDAR, radar, cameras), sensor fusion modules, Electronic Control Units, the internal Controller Area Network, in-vehicle infotainment systems, telematics control units, and Vehicle-to-Everything communication systems. All complex software systems contain bugs, and the extensive codebase of CAVs is no exception; a fraction of these bugs can be exploited as security vulnerabilities. Recent automotive security reports show that such vulnerabilities continue to affect connected vehicle services, with 2024 Kia vulnerabilities enabling remote vehicle control using only a license plate and 2025 Subaru Md Erfan, Ahmed Ryan, and Md Rayhanur Rahman are with the Department of Computer Science, The University of Alabama, Tuscaloosa, USA. Email: {merfan, aryan9}@crimson.ua.edu, [email protected] Md Kamal Hossain Chowdhury is with the Alabama Water Institute, The University of Alabama, Tuscaloosa, USA. Email: [email protected] † Corresponding author.
Fig. 1: An Example of CVE Description to STIX Generation. STARLINK flaws exposing vehicle controls, customer data, and historical location records [3, 4]. The information about these vulnerabilities is provided as plain text, which we refer to as Common Vulnerabilities and Exposures (CVE) descriptions. As a result, security practitioners must analyze unstructured CVE descriptions to identify affected assets, weakness types, attack behaviors, and defensive actions. Converting CVE descriptions into a structured knowledge representation, such as Structured Threat Information Expression (STIX) [18], can organize vulnerability knowledge into entities, relationships, weakness mappings, and attack-behavior mappings, as shown in Fig. 1. Such structured representation helps practitioners compare vulnerabilities, analyze attack paths, and prioritize defenses more systematically. Large Language Models (LLMs) provide languageunderstanding and structured knowledge-generation capabilities, which make them relevant for CVE-to-STIX generation. However, prior studies on STIX generation from cybersecurityrelated text show that LLMs-generated outputs can include incomplete objects, incorrect relationships, invalid structures, and hallucinated attribute values [13, 17]. Therefore, a groundtruth dataset is necessary before LLM-generated STIX can be reliably used for security analysis. Evaluating LLMs requires a ground-truth dataset that connects CVE descriptions with STIX entities, relationships, Common Weakness Enumeration (CWE), and MITRE ATT&CK mappings. Existing datasets for STIX
generation [1, 5, 9] focus on generating MITRE ATT&CK techniques and CWE identification from cybersecurity-related text. However, to the best of our knowledge, no CAV-focused dataset exists that maps CVE descriptions to STIX bundle generation. The goal of this research is to help security practitioners conduct transportation-domain threat analysis by evaluating LLMs for CVE-to-STIX generation in CAV. We address the following research questions (RQ). RQ1 [Dataset Construction]: How can we construct a dataset of STIX representations from unstructured CVE descriptions in the CAV domain? • RQ2 [Open-Weight LLM Evaluation]: How effectively do open-weight LLMs perform in CVE-to-STIX generation under different prompting strategies? • RQ3 [CWE and MITRE ATT&CK Technique Analysis]: What are the frequent CWE and MITRE ATT&CK techniques in CAV-related vulnerabilities, and which ATT&CK techniques co-occur in the dataset?
•
We construct a CAV-focused dataset by manually annotating STIX entities, relationships, CWE mappings, and MITRE ATT&CK technique mappings, and then converting the validated annotations into ground-truth STIX JSON files. Using this dataset, we evaluate open-weight LLMs under multiple prompting strategies and measure performance across STIX Domain Objects (SDO) extraction, STIX Relationship Objects (SRO) extraction, CWE, and MITRE ATT&CK mapping. We also analyze ATT&CK technique co-occurrence in the groundtruth data to identify recurring attack-behavior patterns. We make the following contributions: We construct a CAV-focused benchmark dataset by manually annotating CAV-related CVEs and converting the validated annotations into ground-truth STIX JSON files. • We evaluate open-weight LLMs for generating structured STIX representations from unstructured CVE descriptions. • We compare prompting strategies to examine how different levels of task guidance affect generation quality. • We analyze MITRE ATT&CK techniques in the groundtruth data to identify frequent techniques and recurring co-occurrence patterns in CAV-related vulnerabilities. •
We make the dataset, prompts, and replication scripts available to independent researchers and practitioners (https://figshare.com/s/29f4ceff6bb1c6de5700). The remainder of this paper is organized as follows. Section II introduces the key concepts used in this study, including CVE, CWE, MITRE ATT&CK, and STIX. Section III reviews prior work on cyber threat intelligence datasets, LLM-based STIX generation, and CWE and MITRE ATT&CK technique analysis. Section IV describes the construction of the CAVSTIXGen dataset, the LLM evaluation design, the CWE and MITRE ATT&CK analysis, and the experimental setup. Section V presents the findings for the research questions, including dataset characteristics, open-weight LLM performance, and recurring weakness and attack-behavior patterns in CAV-related
vulnerabilities. Finally, Section VII concludes the paper and discusses future research directions. II. K EY C ONCEPTS In this section, we discuss key concepts related to this study. Common Vulnerabilities and Exposures (CVE) is a public database for known vulnerabilities. Each CVE record provides a unique identifier, such as CVE-2023-289011 , and a naturallanguage description of the vulnerability. CVE records can be found through the official CVE website2 and through vulnerability databases such as the NVD1 . Common Weakness Enumeration (CWE) is a standardized classification of software and hardware weakness types. CWE mappings help identify the root weakness behind a vulnerability, such as improper access control, buffer overflow, injection, or improper input validation. MITRE ATT&CK is a taxonomy for adversary behavior into Tactics, Techniques, and Procedures (TTPs). Tactics describe the attacker’s objective, such as initial access, execution, persistence, privilege escalation, or impact. Techniques describe the methods used to achieve those objectives, such as exploiting a public-facing application or causing endpoint denial of service. MITRE ATT&CK helps security practitioners describe attacks in a consistent vocabulary and connect vulnerability information with known adversary behaviors. Structured Threat Information Expression (STIX) is a standardized format for representing and sharing cyberthreat information. Within STIX, cyber threat knowledge is organized through SDO and SRO. We present a motivational example of converting CVE-2023-289011 into STIX in Fig. 1. The CVE description states that a broken access-control vulnerability in the Skoda Automotive cloud allows remote attackers to access Skoda Connect user information through an arbitrary vehicle Vehicle Identification Number (VIN). The STIX representation separates the plain-English description into entities, including the vendor, cloud infrastructure, affected service, vulnerability, attack pattern, indicator, and remote attacker. The relationships connect these entities by showing that the vulnerability affects the Skoda Connect service, the remote attacker exploits the vulnerability, and the attack pattern targets the affected service. The example also adds CWE-200 for sensitive information exposure and MITRE ATT&CK technique T1190 for exploiting a public-facing application.This structured representation helps security practitioners identify the affected asset, weakness type, attack behavior, and exposed information more efficiently. In STIX, the SDO and SRO are packaged as a JSON-based STIX bundle for machine-readable storage, sharing, and analysis. III. R ELATED W ORK We organize related work according to the main research directions of our study: dataset construction for CTI and vulnerability understanding, LLM-based STIX and CTI generation, and CWE and MITRE ATT&CK pattern analysis. 1 https://nvd.nist.gov/vuln/detail/CVE-2023-28901 2 https://www.cve.org/
Dataset Construction for CTI and Vulnerability Understanding. Prior work has developed various datasets and benchmarks for cyber threat intelligence, vulnerability understanding, and security knowledge extraction. For instance, CTIBench [1] provides a benchmark for evaluating LLMs on CTI tasks, including CWE prediction, MITRE ATT&CK technique extraction, threat actor attribution, and CTI concept understanding. Focusing on vulnerabilities, CVE-LLM [5] uses an ontologyassisted LLM approach for asset and vulnerability based evaluations, while Text2Weak [16] maps CVE descriptions to CWE categories using description embeddings and LLMs. From a structural perspective, STIXnet [9] constructs a pipeline for extracting STIX objects from CTI reports using rulebased methods, knowledge bases, Natural Language Processing (NLP), machine learning, and deep learning. Similarly, CTIGEN [13] generates STIX 2.1-compliant CTI from textual reports and manually constructed attack scenarios. Notably, A recent survey on TTPs extraction reports that vulnerability databases are used less frequently than benchmark datasets, public knowledge bases, and CTI reports [19]. This observation motivates domain-specific CVE-based datasets for evaluating LLMs in structured threat intelligence generation. LLM-Based STIX and CTI Generation. Prior survey work shows that Cyber Threat Intelligence (CTI) extraction studies export structured threat knowledge in multiple formats, including knowledge graphs, STIX, Malware Information Sharing Platform, Open Indicators of Compromise, and domainagnostic representations [14]. Among these formats, STIX is widely used for representing CTI objects and relationships in a machine-readable form, and prior studies have explored how LLMs and Natural Language Processing (NLP) pipelines can generate or extract structured CTI from unstructured sources. For example, Time for aCTIon [17] evaluates GPT-3.5-turbo for converting real-world CTI reports into structured STIX information. Similarly, CTI-GEN [13] uses GPT-4o with multistage prompt engineering to generate STIX 2.1 objects and relationships from textual reports and manually constructed attack scenarios. In contrast, STIXnet [9] focuses on SDO through a hybrid pipeline rather than LLM evaluation. More recently, MAD-CTI [15] evaluates GPT-4o mini, GPT-4o, and Claude 3.5 in a multi-agent framework for dark-web CTI analysis. Together, these studies provide evidence that LLMs and extraction can support structured CTI generation, but CAVrelated CVE descriptions to STIX generation with open-weight LLMs evaluation remains underexplored. CWE and MITRE ATT&CK Technique Analysis. Prior work also examines CWE classification and Mitre attack technique extraction from vulnerability or threat descriptions. For CWE-oriented analysis, CTIBench [1] includes CVE-toCWE prediction and CTI-based MITRE ATT&CK technique extraction tasks. Similarly, Text2Weak [16] focuses on mapping CVE descriptions to CWE categories. In the autonomous vehicle domain, Haque et al. [6] analyze autonomous vehicle software stacks and identify recurring CWE patterns and thirdparty library risks in platforms such as Autoware, Apollo, and openpilot. In parallel, several studies focus on MITRE
ATT&CK and TTP extraction from unstructured CTI. For example, TTPDrill [7] extracts threat actions from unstructured CTI sources and maps them to tactics, techniques, and killchain phases. ThreatPilot [20] extracts Tactics, Techniques, and Procedures from CTI reports and supports Sigma rule generation. TIEF [8] extracts TTPs from threat reports and represents attack behaviors in STIX format. Together, these studies support CWE prediction and ATT&CK/TTP extraction, but they do not focus on CWE and MITRE ATT&CK cooccurrence within CAV-focused ground-truth STIX data. Prior work provides CTI extraction pipelines, LLM-based STIX generation methods, CVE understanding datasets, and ATT&CK/TTP extraction techniques. Our work differs by focusing on CAV-related CVE descriptions, constructing a CAV-focused CVE-to-STIX dataset, evaluating open-weight LLMs for SDO and SRO generation, and analyzing CWE and MITRE ATT&CK mapping in ground-truth STIX data. IV. M ETHODOLOGY We discuss the methodology in this section. A. CAV-STIXGen Dataset Construction We construct the CAV-STIXGen dataset to address RQ1. Step 1: Keyword construction. We built a domain-specific keyword to collect CAV-related vulnerabilities, given in Table I by asking ChatGPT, Copilot, Gemini, and PerplexityChatGPT3 , Copilot4 , Gemini5 , and Perplexity6 . The first and third authors then review, merge, and filter the generated keywords to remove duplicate and weakly related keywords. TABLE I: Unique Keyword Categories Category Types Manufacturers Hardware and Systems Connectivity
Unique Keywords Connected vehicle, Automotive, Electric Vehicle (EV) Audi, Tesla, BMW, Toyota, Mercedes-Benz, Ford, General Motors, Hyundai, Kia, Nissan, Volkswagen, BYD GPS in vehicle, Inertial Measurement Unit (IMU), Steering, Throttle, Controller Area Network (CAN) in vehicle Vehicle-to-Vehicle (V2V), Vehicle-to-Cloud (V2C), Over-the-Air (OTA) updates
Step 2: CVE Collection from NVD. We searched with the keywords the National Vulnerability Database (NVD) through REST API calls. The search returns 445 candidate CVE records. We then remove duplicates and retain 368 records. Step 3: Relevance filtering. The first and second authors assess each CVE record using inclusion and exclusion criteria as follows. We include CVEs related to vehicle components, EV charging, telematics, infotainment, Electronic Control Units (ECUs), CAN, Vehicle-to-Everything (V2X), sensors, and automotive infrastructure. We exclude CVEs with vulnerabilities affecting generic websites, sales platforms, administrative systems, or non-vehicle-specific software. The annotators achieve Cohen’s κ [2] 94% inter-rater agreement, and after resolving disagreements, we retain 183 unique records. 3 https://chatgpt.com/ 4 https://copilot.microsoft.com/ 5 https://gemini.google.com/app 6 https://www.perplexity.ai/
Step 4: STIX Domain Objects and STIX Relationship Objects Annotation. We manually annotate SDO and SRO for each retained CVE according to the STIX 2.1 specification [12]. We store the annotations in CSV files to enable review, correction, and validation. The annotations include the relevant STIX object types among the 19 SDO types. We then annotate SRO by linking each source object to a target object using the relevant relationship types. We assign each object and relationship a unique identifier, denoted as Ei and Ri , respectively, where i is a natural number. Step 5: CWE and MITRE ATT&CK Mapping. We map each CVE with corresponding CWE and Mitre attack technique mappings. We collect CWE information from the NVD when available. We then manually verify each NVDprovided CWE mapping against the corresponding CVE description to ensure that the assigned weakness type matches the vulnerability behavior. When the NVD does not provide a CWE mapping, such as for CVE-2024-510747 , we manually analyze the CVE description and assign the appropriate CWE based on the vulnerability behavior and root cause. We also map each extracted attack-pattern STIX object to the corresponding MITRE ATT&CK techniques [11]. Step 6: Validation and STIX JSON Conversion. After Fig. 2: Prompting strategies for CVE-to-STIX generation. completing Step 5, we store all annotations in CSV files for inspection and correction. The first and second authors review 4-31B, Microsoft Phi-4, and Qwen-3.5-9B. Cybersecuritythe files, resolve disagreements, and revise ambiguous cases. specialized models: Lily-CyberSecurity-7B and CyberSecAfter validation, we programmatically convert the CSV files Qwen3-DeepSeekv1. Code-oriented models: Codestral-22B into ground-truth STIX 2.1 JSON bundles. and Qwen3-Coder-30B. Large-capacity open-weight models: GPT-OSS-20B, GPT-OSS-120B, and LLaMA-3.3-70B. We B. LLM Evaluation Using the CAV-STIXGen Dataset evaluate a multi-agent configuration with Gemma-4-31B and We use the CAV-STIXGen dataset to evaluate LLM-based Codestral to examine whether role-specific decomposition CVE-to-STIX generation. Each model receives a CAV-related improves generation compared with single-model prompting. CVE identifier and description as input and generates a STIX For space, tables use shortened model names where needed.8 JSON output. We compare each generated output with the manFig. 2 summarizes the three prompting strategies used in our ually validated ground-truth STIX file for the same CVE. We evaluation. Contextless prompting uses the CVE identifier and evaluate each model under multiple prompting strategies and vulnerability description as inputs for generating a STIX JSON four temperature settings: 0, 0.25, 0.75, and 1.0, representing bundle. STIX-guided prompting uses the same input and deterministic generation, low-variation generation, moderate- adds task guidance for STIX object types, CWE representation, variation generation, and high-variation generation, respectively. MITRE ATT&CK mapping, and relationship construction. The evaluation measures SDO extraction, SRO construction, Dynamic few-shot prompting extends the guided setting with CWE mapping, and MITRE ATT&CK technique mapping. We five similar ground-truth STIX examples using similarity score use micro-precision, micro-recall, and micro-F1 because CVE- among description, CWE, and attack-pattern. to-STIX generation is an instance-level generation task, and micro-level aggregation captures overall correctness across all C. CWE and MITRE ATT&CK Technique Analysis generated objects, relationships, CWE, and MITRE ATT&CK We perform the CWE and MITRE ATT&CK Technique mappings without over-weighting rare labels. For MITRE Analysis to address RQ3. ATT&CK mapping, we also report Match@1 and Match@All For CWE analysis, we examine the weakness mappings because a single CVE can be associated with multiple valid associated with each CAV-related CVE in the ground-truth data. ATT&CK techniques. Match@1 measures whether a model We first count the frequency of each CWE category across the identifies at least one correct technique, while Match@All dataset to identify dominant weakness patterns in CAV-related measures whether the model identifies all techniques from the vulnerabilities. We then compare the frequent CWE categories ground-truth for that CVE. 8 Note: For space, Microsoft Phi-4 is shown as Phi-4, Qwen-3-CoderWe evaluate 11 open-weight LLMs across four model 30B as Qwen-Coder, Qwen-3.5-9B as Qwen-9B GPT-OSS-20B as GPT-20B, categories. General-purpose models: Gemma-3-4B, Gemma- GPT-OSS-120B as GPT-120B, and LLaMA-3.3-70B as LLaMA-70B, Lily7 https://nvd.nist.gov/vuln/detail/CVE-2024-51074
CyberSecurity-7B as Lily-7B, CyberSec-Qwen3-DeepSeekv1 as CySec-Qwen and Codestral-22B as Codestral-22B.
with the CWE Top 25 list [10] to contextualize whether the and attacker roles. The most frequent SRO types are observed weaknesses align with widely recognized software exploits, targets, uses, affects, part-of, security risks. This analysis helps characterize the weakness produces, enables, has-version, deployed-in, landscape of CAV vulnerabilities and explains which types of and used-by. These relationship types show how attack software flaws appear most often in the dataset. behaviors exploit vulnerabilities, target systems, use resources, Although MITRE ATT&CK techniques are represented affect infrastructures, and connect software components to in STIX through attack-pattern SDOs, we analyze CAV-related assets. On average, each CVE contains 7.56 SDO ATT&CK mappings separately to measure whether these instances, 7.62 SRO instances, 1.15 CWE mappings, and attack-pattern objects are correctly normalized to 1.61 MITRE ATT&CK mappings. MITRE technique IDs. We also perform MITRE ATT&CK technique co-occurrence analysis to identify recurring attack- B. RQ2: Open-Weight LLM Evaluation behavior patterns in the ground-truth STIX data. Each CVE RQ2 evaluates how effectively open-weight LLMs generate is treated as one transaction, and the MITRE ATT&CK STIX representations from CAV-related CVE descriptions under techniques extracted from the attack-pattern STIX obdifferent prompting strategies, shown in Table II. We compare jects are treated as transaction items. We extract MITRE contextless prompting, STIX-guided prompting, and dynamic technique IDs from the external_references field of few-shot prompting across five evaluation dimensions: STIX each attack-pattern object. We then apply frequent object extraction, STIX relationship extraction, CWE mapping, itemset mining using FP-Growth with a minimum support MITRE ATT&CK Match@1, and Match@All. threshold of 0.02 to identify frequent individual techniques 1) Overall Prompting Strategy Comparison: The results and technique combinations. We also generate association show a performance difference across prompting strategies. rules using a minimum confidence threshold of 0.30 to Contextless prompting provides weaker performance because measure directional associations between techniques. Finally, models receive only minimal task instructions and infer STIX we construct a pairwise co-occurrence network from the without examples. STIX-guided prompting improves STIX extracted techniques, where each node represents a MITRE generation by providing explicit instructions for SDO types, technique, node size represents technique frequency, and edge weight represents the number of CVEs in which two techniques SRO, CWE references, and MITRE ATT&CK mappings. Dynamic few-shot prompting gives the strongest overall results appear together. because target-relevant examples help models produce more D. Experimental Setup complete STIX bundles. For example, Phi-4 achieves an SDO We conducted the CVE-to-STIX generation experiments extraction F1 score of 0.94 under dynamic few-shot, while using a local inference server and a development machine. The Qwen-Coder achieves a relationship extraction F1 score of inference server ran LM Studio version 0.4.8 (Build 1) on 0.63. The results indicate that SDO examples improves LLMs Ubuntu and hosted the open-weight LLMs through the LM performance on STIX generation, while dynamically selected Studio OpenAI-compatible API. The server used four NVIDIA examples improve both structure and semantic mapping. 2) Performance Across CVE-to-STIX Tasks: In this task, RTX 6000 Ada Generation GPUs with 48 GB of VRAM STIX object extraction obtains the highest scores across most each, totaling 192 GB of GPU memory, with NVIDIA Driver models. Under dynamic few-shot prompting, Phi-4, LLaMAversion 580.126.09 and CUDA 13.0. 70B, Codestral-22B, and Qwen-Coder reach object extraction V. R ESULTS F1 scores around 0.94. Relationship extraction obtains lower scores because the task requires models to identify the correct We present the results for RQ1-RQ3 in this section. source object, target object, and relationship type. Qwen-Coder A. RQ1: Construction of CAV-STIXGen Dataset achieves the best relationship extraction result, with an F1 Ground-Truth STIX Bundle Characteristics. Table IV score of 0.63 under dynamic few-shot prompting. CWE mapping shows improvement when similar examples summarizes the composition of CAV-STIXGen. The dataset starts from 445 candidate CVE records collected from are provided. Under dynamic few-shot prompting, CodestralNVD and retains 183 CAV-related CVEs after relevance 22B, Qwen-Coder, LLaMA-70B, and GPT-120B each achieve filtering. The final ground-truth dataset contains 1,383 SDO CWE F1 scores close to 0.99. This result indicates that instances, 1,395 SRO instances, 211 CWE mappings, and example-guided prompting helps models connect vulnerability 294 MITRE ATT&CK mappings. The dataset covers 10 out descriptions with weakness categories and normalize CWE of 19 SDO types: infrastructure, attack-pattern, identifiers more accurately than contextless prompting. vulnerability, identity, threat-actor, note, MITRE ATT&CK mapping requires richer attack-behavior tool, indicator, observed-data, and malware. reasoning than CWE mapping. Match@1 improves under The most frequent SDO types are infrastructure, dynamic few-shot prompting, which indicates that models attack-pattern, vulnerability, identity, and can often identify at least one correct ATT&CK technique. threat-actor, indicating that the dataset captures Match@All remains lower because a single CVE can involve affected assets, attack behaviors, weaknesses, vendors, multiple valid ATT&CK techniques.
TABLE II: CVE-to-STIX generation performance across prompting strategies Model
T
Gemma-3-4B
Gemma-4-31B
Phi-4
Qwen-9B
Lily-7B
CySec-Qwen
Codestral-22B
Qwen-Coder
GPT-20B
LLaMA-70B
GPT-120B
0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00 0.00 0.25 0.75 1.00
SDO SG 0.84 0.85 0.83 0.77 0.86 0.86 0.88 0.88 0.89 0.89 0.87 0.86 0.14 0.17 0.15 0.11 0.83 0.78 0.69 0.62 0.28 0.20 0.08 0.06 0.89 0.88 0.88 0.86 0.84 0.85 0.84 0.84 0.85 0.85 0.87 0.84 0.88 0.88 0.88 0.88 0.90 0.90 0.90 0.90
C 0.60 0.59 0.53 0.50 0.60 0.60 0.60 0.60 0.65 0.65 0.60 0.59 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.52 0.52 0.41 0.32 0.57 0.56 0.49 0.48 0.54 0.55 0.54 0.55 0.60 0.61 0.60 0.57 0.68 0.69 0.66 0.67 0.59 0.60 0.60 0.60
DFS 0.91 0.91 0.90 0.89 0.79 0.80 0.76 0.78 0.94 0.94 0.94 0.94 0.80 0.75 0.75 0.78 0.87 0.87 0.83 0.78 0.86 0.84 0.71 0.61 0.94 0.93 0.92 0.88 0.94 0.94 0.93 0.94 0.84 0.81 0.77 0.75 0.94 0.94 0.94 0.94 0.83 0.83 0.82 0.83
C 0.00 0.00 0.00 0.00 0.20 0.20 0.20 0.21 0.01 0.02 0.03 0.03 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.15 0.15 0.13 0.13 0.03 0.04 0.04 0.04 0.10 0.12 0.11 0.09 0.13 0.12 0.13 0.12 0.20 0.18 0.17 0.18
SRO SG 0.03 0.03 0.03 0.01 0.45 0.44 0.46 0.47 0.40 0.38 0.38 0.35 0.07 0.08 0.05 0.04 0.00 0.00 0.00 0.00 0.15 0.09 0.03 0.02 0.22 0.20 0.17 0.17 0.38 0.36 0.37 0.37 0.41 0.40 0.39 0.37 0.38 0.38 0.39 0.38 0.49 0.48 0.48 0.47
DFS 0.31 0.31 0.30 0.28 0.60 0.61 0.60 0.61 0.55 0.55 0.53 0.53 0.50 0.47 0.44 0.46 0.37 0.41 0.31 0.25 0.39 0.38 0.28 0.20 0.55 0.54 0.53 0.50 0.62 0.62 0.62 0.63 0.51 0.48 0.46 0.45 0.51 0.51 0.51 0.50 0.52 0.53 0.52 0.51
C 0.00 0.00 0.00 0.00 0.39 0.38 0.38 0.37 0.64 0.65 0.58 0.56 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.11 0.10 0.12 0.07 0.01 0.02 0.01 0.03 0.12 0.12 0.11 0.12 0.09 0.10 0.13 0.13 0.31 0.28 0.30 0.27
CWE SG 0.11 0.11 0.11 0.10 0.39 0.37 0.38 0.37 0.89 0.89 0.85 0.84 0.08 0.09 0.07 0.09 0.11 0.10 0.10 0.11 0.05 0.02 0.01 0.02 0.16 0.15 0.18 0.19 0.03 0.01 0.02 0.01 0.27 0.25 0.23 0.17 0.35 0.35 0.31 0.31 0.35 0.33 0.29 0.30
DFS 0.95 0.95 0.95 0.93 0.93 0.95 0.92 0.94 0.94 0.94 0.93 0.94 0.81 0.76 0.77 0.80 0.96 0.95 0.90 0.82 0.90 0.89 0.73 0.61 0.99 0.98 0.98 0.96 0.99 0.99 0.98 0.99 0.95 0.93 0.88 0.85 0.99 0.99 0.98 0.99 0.98 0.99 0.99 0.98
C 0.02 0.01 0.02 0.01 0.50 0.52 0.47 0.43 0.14 0.12 0.09 0.06 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.02 0.02 0.01 0.15 0.11 0.07 0.08 0.23 0.20 0.19 0.17 0.17 0.20 0.21 0.18 0.17 0.24 0.17 0.15 0.34 0.39 0.39 0.36
Attack Match@1 SG 0.17 0.16 0.17 0.14 0.41 0.41 0.42 0.42 0.23 0.22 0.15 0.18 0.04 0.04 0.05 0.02 0.07 0.07 0.02 0.05 0.09 0.07 0.01 0.01 0.29 0.26 0.22 0.21 0.27 0.28 0.23 0.25 0.28 0.26 0.23 0.20 0.12 0.14 0.13 0.14 0.44 0.37 0.39 0.41
DFS 0.37 0.39 0.38 0.37 0.64 0.68 0.62 0.66 0.53 0.51 0.52 0.49 0.52 0.45 0.47 0.49 0.28 0.29 0.20 0.13 0.42 0.39 0.24 0.14 0.46 0.45 0.46 0.42 0.67 0.64 0.64 0.67 0.56 0.54 0.46 0.49 0.61 0.61 0.59 0.57 0.57 0.58 0.61 0.57
C 0.02 0.01 0.02 0.01 0.24 0.23 0.24 0.22 0.03 0.03 0.01 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.01 0.01 0.00 0.07 0.04 0.03 0.02 0.11 0.10 0.09 0.09 0.07 0.09 0.07 0.08 0.08 0.12 0.09 0.08 0.15 0.17 0.16 0.14
Attack Match@All SG DFS 0.04 0.16 0.04 0.16 0.04 0.15 0.03 0.14 0.19 0.28 0.18 0.33 0.20 0.28 0.18 0.33 0.08 0.22 0.09 0.22 0.07 0.20 0.04 0.19 0.01 0.23 0.02 0.22 0.02 0.23 0.01 0.22 0.04 0.12 0.03 0.11 0.00 0.08 0.02 0.05 0.02 0.19 0.02 0.19 0.00 0.10 0.01 0.09 0.14 0.19 0.10 0.19 0.09 0.19 0.09 0.18 0.12 0.26 0.13 0.27 0.08 0.27 0.13 0.28 0.13 0.29 0.11 0.25 0.10 0.24 0.07 0.24 0.03 0.27 0.06 0.26 0.06 0.27 0.06 0.26 0.20 0.27 0.18 0.29 0.17 0.30 0.18 0.27
Note: C = Contextless, SG = STIX-Guided, DFS = Dynamic Few-Shot. Objects, Relationship, and CWE are F1 scores. Match@1 and Match@All are MITRE ATT&CK mapping scores. Green box represents the largest value in a column.
TABLE III: Multi-Agent SDO and SRO Extraction Model
Temp.
MultiAgent (Gemma-4-31B) MultiAgent (Codestral-22B)
0.75 0
Objects P R F1 0.94 0.88 0.91 0.90 0.83 0.86
Relationship P R F1 0.38 0.40 0.39 0.50 0.38 0.43
TABLE IV: Statistics of CAV-STIXGen Dataset Composition Dataset Characteristic Candidate CVE records collected from NVD Retained CAV-related CVEs Data collection period Most frequent CVE year Total STIX objects Total STIX relationships Average STIX objects per CVE Average STIX relationships per CVE STIX object types covered Most frequent STIX object type Total CWE mappings Average CWE mappings per CVE Most frequent CWE Top-10 CWEs appearing in CWE Top 25 Total MITRE ATT&CK mappings Average MITRE ATT&CK mappings per CVE Most frequent MITRE ATT&CK technique Most frequent ATT&CK co-occurring pair Annotation format
Value 445 183 2012–2025 2023, 61 CVEs 1,383 1,395 7.56 7.62 10 of 19 Infrastructure, 538 211 1.15 CWE-787, CWE-20 5 of 10 294 1.61 T1499, 48 T1203–T1499, 10 CVEs CSV & JSON
3) Model-Level Findings: Model performance varies across subtasks. LLaMA-70B, Phi-4, Codestral-22B, Qwen-Coder, and GPT-120B perform strongly for STIX object extraction under guided or few-shot prompting. Qwen-Coder achieves the strongest relationship extraction score, followed by Gemma4-31B, Codestral-22B, Phi-4, and GPT-120B under dynamic
few-shot prompting. For CWE mapping, dynamic few-shot prompting produces the most consistent gains across larger and code-oriented models. For MITRE ATT&CK mapping, Gemma4-31B achieves the highest Match@1 score of 0.68, followed by Qwen-Coder with 0.67. These results suggest that no single model is best for every subtask; object extraction, relationship construction, CWE mapping, and ATT&CK mapping require different reasoning capabilities. 4) SDO Type Analysis: Table V shows that performance differs across STIX object types. Vulnerability extraction is consistently strong because CVE descriptions explicitly describe the vulnerability. For example, Phi-4, Codestral-22B, QwenCoder, and LLaMA-70B reach vulnerability F1 scores close to 0.99 in at least one prompting setting. Other object types require more contextual inference. Threat-actor, infrastructure, and identity objects are missed under contextless prompting, but their extraction improves with STIX-guided or dynamic few-shot prompting. Dynamic few-shot prompting is especially effective for implicit object types because similar examples show how contextual information should be represented as STIX objects. 5) STIX Relationship Object Type Analysis: Table VI reports F1 scores for the top 10 SRO types across all evaluated LLMs. The analysis focuses on these frequent relationship types because they represent the dominant edge patterns in the ground-truth STIX bundles. Relationship extraction is
TABLE V: STIX Domain Object Type F1 Scores Model
Prompt C (0) Gemma-3-4B SG (0.25) DFS (0) C (0) Gemma-4-31B SG (1.0) DFS (0.25) C (0) Phi-4 SG (0.25) DFS (0.25) SG (0.25) Qwen-9B DFS (0) SG (0) Lily-7B DFS (0.25) C (0.25) CySec-Qwen SG (0) DFS (0) C (0) Codestral-22B SG (0) DFS (0) C (1.0) Qwen-Coder SG (0.25) DFS (0.25) C (0.25) GPT-20B SG (0.75) DFS (0) C (0.25) LLaMA-70B SG (0.25) DFS (1.0) C (0.25) GPT-120B SG (0) DFS (1.0)
AP 0.65 0.78 0.75 0.75 0.80 0.83 0.72 0.76 0.78 0.20 0.67 0.74 0.55 0.67 0.35 0.73 0.70 0.80 0.81 0.74 0.77 0.87 0.65 0.77 0.77 0.74 0.78 0.79 0.73 0.81 0.79
Vul. 0.88 0.98 0.98 0.99 0.96 0.95 0.98 0.99 0.99 0.30 0.83 0.99 0.98 0.94 0.50 0.94 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.98 0.95 0.99 0.99 0.99 0.99 0.99 0.99
TA 0.53 0.82 0.82 0.00 0.82 0.89 0.04 0.62 0.91 0.08 0.80 0.59 0.84 0.00 0.04 0.67 0.00 0.75 0.92 0.00 0.21 0.90 0.00 0.64 0.85 0.00 0.58 0.89 0.00 0.84 0.88
ID Infra. 0.34 0.05 0.42 0.59 0.84 0.68 0.14 0.03 0.77 0.62 0.80 0.83 0.61 0.00 0.82 0.49 0.82 0.73 0.26 0.11 0.69 0.71 0.62 0.34 0.81 0.73 0.01 0.00 0.34 0.21 0.78 0.57 0.00 0.00 0.80 0.46 0.84 0.75 0.21 0.00 0.79 0.53 0.84 0.81 0.56 0.00 0.81 0.56 0.79 0.78 0.74 0.00 0.79 0.51 0.82 0.76 0.22 0.00 0.79 0.54 0.82 0.82
Note: AP = Attack Pattern, Vul. = Vulnerability, TA = Threat Actor, ID = Identity, Infra. = Infrastructure, C = Contextless, SG = STIX-Guided, DFS = Dynamic Few-Shot. Temperature values are shown in parentheses.
TABLE VI: STIX Relationship Object Type F1 Scores Model
Prompt Exp. Tar. Use Aff. Part Prod. Enab. Ver. Dep. U-by SG (0) 0.08 0.02 0.06 0.02 0.00 0.00 0.00 0.00 0.00 0.00 DFS (0) 0.22 0.01 0.09 0.90 0.50 0.08 0.00 0.03 0.18 0.36 C (1.0) 0.56 0.29 0.05 0.07 0.00 0.00 0.00 0.00 0.00 0.00 Gemma-4-31B SG (1.0) 0.70 0.13 0.63 0.86 0.26 0.19 0.39 0.00 0.24 0.00 DFS (1.0) 0.65 0.38 0.65 0.84 0.71 0.72 0.00 0.65 0.42 0.29 C (0.75) 0.00 0.07 0.15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Phi-4 SG (0) 0.28 0.61 0.32 0.89 0.03 0.53 0.00 0.00 0.16 0.00 DFS (0) 0.56 0.44 0.45 0.91 0.58 0.68 0.00 0.43 0.44 0.10 SG (0.25) 0.08 0.09 0.05 0.28 0.00 0.02 0.06 0.00 0.00 0.00 Qwen-9B DFS (0) 0.56 0.31 0.56 0.75 0.53 0.56 0.00 0.61 0.21 0.00 Lily-7B DFS (0.25) 0.34 0.29 0.16 0.82 0.62 0.24 0.00 0.40 0.28 0.33 C (0) 0.00 0.00 0.05 0.00 0.00 0.00 0.00 0.00 0.00 0.00 CySec-Qwen SG (0) 0.30 0.38 0.09 0.02 0.01 0.00 0.00 0.00 0.00 0.00 DFS (0) 0.42 0.18 0.18 0.85 0.50 0.32 0.00 0.15 0.18 0.00 C (0) 0.01 0.15 0.63 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Codestral-22B SG (0) 0.62 0.42 0.52 0.77 0.01 0.33 0.02 0.00 0.00 0.00 DFS (0) 0.58 0.39 0.65 0.92 0.53 0.58 0.00 0.34 0.35 0.12 C (1.0) 0.14 0.07 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Qwen-Coder SG (0) 0.67 0.25 0.07 0.74 0.07 0.33 0.07 0.00 0.00 0.00 DFS (1.0) 0.68 0.60 0.60 0.91 0.68 0.72 0.02 0.54 0.49 0.32 C (0.25) 0.01 0.34 0.34 0.04 0.00 0.00 0.00 0.00 0.00 0.00 GPT-20B SG (0) 0.59 0.43 0.56 0.83 0.09 0.08 0.00 0.03 0.11 0.00 DFS (0) 0.52 0.42 0.65 0.83 0.61 0.28 0.00 0.43 0.38 0.36 C (0.75) 0.13 0.50 0.09 0.01 0.00 0.00 0.00 0.00 0.00 0.00 LLaMA-70B SG (0.75) 0.66 0.14 0.37 0.91 0.01 0.12 0.00 0.00 0.00 0.00 DFS (0.75) 0.55 0.28 0.40 0.91 0.66 0.45 0.00 0.25 0.27 0.33 C (0) 0.27 0.27 0.53 0.01 0.00 0.00 0.00 0.00 0.00 0.00 GPT-120B SG (0) 0.67 0.30 0.62 0.90 0.15 0.63 0.07 0.00 0.00 0.00 DFS (0.25) 0.53 0.48 0.67 0.89 0.66 0.54 0.00 0.53 0.41 0.25 Gemma-3-4B
Note: Exp. = exploits, Tar. = targets, Use = uses, Aff. = affects, Part = part-of, Prod. = produces, Enab. = enables, Ver. = has-version, Dep. = deployed-in, U-by = used-by.
Gemma-4-31B and Codestral-22B multi-agent configuration. We select Gemma-4-31B and Codestral-22B for the multiagent setting because their single-prompt results show strong performance across SDO extraction, relationship construction, and STIX generation. The results indicate that role-based decomposition can improve SDO extraction for some models, but the improvement does not consistently transfer to SRO extraction. Relationship construction remains more complex because the configuration must connect the correct source and target objects through valid STIX relationship types.
more complex than object extraction because models must identify the relevant STIX Domain Objects, determine the C. RQ3: CWE and MITRE ATT&CK Technique Analysis correct source and target objects, and assign the appropriate relationship type. Under contextless prompting, model outputs Table VII shows that CWE-787, CWE-284, and CWE-20 mainly capture explicit relationships such as uses, targets, are the most frequent weaknesses in our dataset. Five of the and exploits. In contrast, semantically richer relationships top 10 CWEs also appear in the 2025 CWE Top 25 [10] list: such as affects, part-of, produces, has-version, CWE-787, CWE-284, CWE-20, CWE-476, and CWE-863. deployed-in, and used-by often receive low F1 scores. These weaknesses cover 57 of the 94 top-10 CWE occurrences, STIX-guided prompting improves relationship extraction by or 60.64%. The overlap indicates that CAV-related vulneraproviding examples on source-target linking and relationship bilities often involve widely recognized high-risk weakness semantics. Dynamic few-shot prompting provides the broad- categories, especially memory-safety, access-control, inputest relationship-type coverage because similar ground-truth validation, and authorization weaknesses. examples demonstrate how vulnerabilities, attack patterns, Table VIII shows that the top 10 MITRE ATT&CK techinfrastructures, and vendors should be connected in STIX. niques account for 227 out of 294 MITRE ATT&CK mappings, Overall, these results indicate that relationship extraction covering 77.21% of all technique mappings. The most frequent requires graph-level reasoning beyond object identification. techniques are T1499 Endpoint Denial of Service, T1203 6) Multi-Agent STIX Generation: The multi-agent configu- Exploitation for Client Execution, and T1210 Exploitation of ration separates CVE-to-STIX generation into SDO extraction, Remote Services. These results show that denial of service, CWE mapping, MITRE ATT&CK mapping, relationship con- client-side exploitation, and remote-service exploitation are struction, STIX JSON generation, and validation. Table III recurring attack behaviors in CAV-related vulnerabilities. The summarizes the SDO and SRO extraction performance of the CWE and MITRE ATT&CK findings suggest that CWE
TABLE VII: Top 10 CWE Categories in Dataset
Qwen3-Coder-30B achieves the highest relationship extraction F1 score of 0.6298 under dynamic few-shot prompting. This gap shows that CVE-to-STIX generation requires graph-level reasoning beyond entity recognition. Implication 3: CWE and MITRE ATT&CK mappings require different reasoning. CWE mapping improves under dynamic few-shot prompting because similar examples help models connect vulnerability behavior with standardized weakness categories. MITRE ATT&CK mapping remains more demanding because one CVE can correspond to multiple valid attack techniques. Match@1 improves for models such as TABLE VIII: Top 10 ATT&CK Techniques in Our Dataset Gemma4 and Qwen3-Coder-30B, but Match@All remains Rank Technique Name Count lower. These results suggest that future systems should combine 1 T1499 Endpoint Denial of Service 48 LLM generation with external validation from CWE and 2 T1203 Exploitation for Client Execution 42 3 T1210 Exploitation of Remote Services 34 MITRE ATT&CK knowledge sources. 4 T1200 Hardware Additions 19 Implication 4: Multi-agent generation provides task5 T1557 Adversary-in-the-Middle 18 6 T1078 Valid Accounts 14 specific gains. The Gemma4-based multi-agent configuration 7 T1005 Data from Local System 14 performs better for entity extraction, while the Codestral-22B8 T1068 Exploitation for Privilege Escalation 13 9 T1552 Unsecured Credentials 13 based configuration performs better for relationship extraction. 10 T1548 Abuse Elevation Control Mechanism 12 Total top-10 technique occurrences 227 However, multi-agent decomposition does not consistently Coverage over all MITRE mappings 227/294 = 77.21% outperform the strongest dynamic few-shot setting. This result categories often lead to repeated exploitation and impact indicates that role-based decomposition can help selected subtasks, but the benefit depends on the base model and agent patterns in the transportation security domain. The co-occurrence results further show that role. Implication 5: CAV vulnerabilities contain recurring T1203--T1499 is the most frequent technique pair, weakness and attack-behavior patterns. The ground-truth appearing in 10 CVEs with support 0.0556. The next most dataset includes frequent CWE categories that overlap with the frequent pairs are T1210--T1499 with 6 co-occurrences and 2025 CWE Top 25 list, including CWE-787, CWE-284, CWET1068--T1548 with 5 co-occurrences. The association-rule 20, CWE-476, and CWE-863. The MITRE ATT&CK analysis results also show a bidirectional association between T1548 identifies T1499, T1203, and T1210 as the most frequent techAbuse Elevation Control Mechanism and T1068 Exploitation niques, while T1203–T1499 is the most frequent co-occurring for Privilege Escalation, with confidence values of 0.4167 pair. These patterns show that CAV-related vulnerabilities often and 0.3846. The most frequent MITRE ATT&CK techniques involve connected access-control, input-validation, exploitaare T1499, T1203, and T1210. The most common cotion, denial-of-service, and privilege-escalation behaviors. The occurring technique pair is T1203--T1499, followed by dataset-level findings support the use of CAV-STIXGen for T1210--T1499 and T1068--T1548. both LLM evaluation and transportation-domain threat analysis. VI. D ISCUSSION VII. C ONCLUSION AND F UTURE W ORK Our results provide key findings about open-weight LLMCWE Name CWE-787 Out-of-bounds Write CWE-284 Improper Access Control CWE-20 Improper Input Validation CWE-294 Authentication Bypass by Capture-replay CWE-287 Improper Authentication CWE-693 Protection Mechanism Failure CWE-190 Integer Overflow or Wraparound CWE-476 NULL Pointer Dereference CWE-863 Incorrect Authorization CWE-310 Cryptographic Issues Top-10 CWEs appearing in CWE Top 25 Occurrences covered by CWE Top 25
Count 20 14 12 10 9 8 6 6 5 4
CWE Top 25 Yes Yes Yes No No No No Yes Yes No 5/10 57/94 = 60.64%
We presented CAV-STIXGen, a CAV-focused dataset for based CVE-to-STIX generation for CAV vulnerabilities. Implication 1: Prompt guidance improves CVE-to-STIX evaluating LLM-based CVE-to-STIX generation. The dataset generation. Contextless prompting provides a limited baseline maps CAV-related CVE descriptions to validated SDO, SRO, because models receive minimal task instructions. Models can CWE, and MITRE ATT&CK technique mappings. Using this extract directly stated vulnerability objects, but they often dataset, we evaluated 11 open-weight LLMs ranging from 4B miss implicit STIX objects, relationships, CWE mappings, to 120B parameters across prompting strategies, temperature and MITRE ATT&CK mappings. STIX-guided prompting configurations, and multi-agent settings. The best single-model improves structural completeness through explicit instructions, results reached 0.94 F1 for SDO extraction, 0.63 F1 for SRO while dynamic few-shot prompting achieves the highest overall extraction, and 0.99 F1 for CWE mapping, while complete performance in our results. These results indicate that reliable ATT&CK mapping remained more challenging because one CVE-to-STIX generation requires both schema guidance and CVE can involve multiple valid attack techniques. Our results show that prompt guidance improves CVE-to-STIX generation. target-relevant examples. Implication 2: Object extraction performs better than Contextless prompting produces partial STIX representations, relationship extraction. STIX object extraction reaches high STIX-guided prompting improves structural completeness, and performance under guided and example-based prompting, with dynamic few-shot prompting achieves the strongest overall multiple models reaching F1 scores around 0.94. Relationship performance in our evaluation. The multi-agent setting provides extraction remains more complex because models must identify task-specific gains, with Gemma-4-31B achieving 0.91 F1 for the correct source object, target object, and relationship type. SDO extraction and Codestral-22B achieving 0.43 F1 for SRO
extraction. Dataset-level analysis further identifies recurring weakness and attack-behavior patterns in the CAV domain, including overlap with the CWE Top 25 and repeated MITRE ATT&CK co-occurrence patterns. Future work will extend CAV-STIXGen with additional CAVrelated vulnerabilities, richer STIX object types, and more detailed relationship annotations. Future studies can also evaluate retrieval-augmented prompting with external STIX, CWE, and MITRE ATT&CK knowledge bases. Another direction is to improve relationship construction through graph validation, source-target consistency checking, and post-generation repair. Future multi-agent pipelines can assign specialized models to entity extraction, relationship construction, CWE mapping, MITRE ATT&CK mapping, and validation. These extensions can improve the reliability of AI-assisted CVE-to-STIX generation for transportation-domain threat analysis. R EFERENCES [1] Md Tanvirul Alam, Dipkamal Bhusal, Le Nguyen, and Nidhi Rastogi. 2024. Ctibench: A benchmark for evaluating llms in cyber threat intelligence. Advances in Neural Information Processing Systems 37 (2024). [2] Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement 20, 1 (1960), 37–46. [3] Sam Curry. 2024. Hacking Kia: Remotely Controlling Cars With Just a License Plate. https://samcurry.net/ hacking-kia. Accessed: 2026-05-26. [4] Sam Curry and Shubham Shah. 2025. Hacking Subaru: Tracking and Controlling Cars via the STARLINK Admin Panel. https://samcurry.net/hacking-subaru. Accessed: 2026-05-26. [5] Rikhiya Ghosh, Hans-Martin von Stockhausen, Martin Schmitt, George Marica Vasile, Sanjeev Kumar Karn, and Oladimeji Farri. 2025. Cve-llm: Ontology-assisted automatic vulnerability evaluation using large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 28757–28765. [6] Md Wasiul Haque, Md Erfan, Sagar Dasgupta, Md Rayhanur Rahman, and Mizanur Rahman. 2025. Security Vulnerabilities in Software Supply Chain for Autonomous Vehicles. arXiv:2509.16899 (2025). [7] Ghaith Husari, Ehab Al-Shaer, Mohiuddin Ahmed, Bill Chu, and Xi Niu. 2017. Ttpdrill: Automatic and accurate extraction of threat actions from unstructured text of cti sources. In Proceedings of the 33rd annual computer security applications conference. 103–115. [8] Anooja Joy, Madhav Chandane, Yash Nagare, and Faruk Kazi. 2025. Threat Intelligence Extraction Framework (TIEF) for TTP Extraction. Journal of Cybersecurity and Privacy 5, 3 (2025), 63. [9] Francesco Marchiori, Mauro Conti, and Nino Vincenzo Verde. 2023. Stixnet: A novel and modular solution for extracting all stix objects in cti reports. In Proceedings of the 18th international conference on availability, reliability and security. 1–11.
[10] MITRE. 2025. 2025 CWE Top 25 Most Dangerous Software Weaknesses. https://cwe.mitre.org/top25/archive/ 2025/2025 cwe top25.html. Accessed: 2026-05-28. [11] MITRE. 2025. MITRE ATT&CK. https://attack.mitre.org/. Accessed: 2026-05-25. [12] OASIS Open. 2021. STIX Version 2.1. https://www. oasis-open.org/standard/6426/. Approved: 25 January 2021. Accessed: 2026-05-28. [13] Angelos Papoutsis, Athanasios Dimitriadis, Dimitrios Kavallieros, Theodora Tsikrika, Stefanos Vrochidis, Ioannis Kompatsiaris, and Georgios Meditskos. 2025. Cti-gen: A framework for generating stix 2.1 compliant cti using generative ai. In 2025 IEEE International Conference on Cyber Security and Resilience (CSR). IEEE. [14] Md Rayhanur Rahman, Rezvan Mahdavi Hezaveh, and Laurie Williams. 2023. What are the attackers doing now? automating cyberthreat intelligence extraction from text on pace with the changing threat landscape: A survey. Comput. Surveys 55, 12 (2023), 1–36. [15] Sayuj Shah and Vijay K Madisetti. 2025. MAD-CTI: Cyber threat intelligence analysis of the dark web using a multi-agent framework. IEEE Access (2025). [16] Stefano Simonetto, Thijs Sebastiaan van Ede, Peter Bosch, Willem Jonker, and Ronan Oostveen. 2024. Text2Weak: mapping CVEs to CWEs using description embeddings analysis. In 4th Workshop on Artificial IntelligenceEnabled Cybersecurity Analytics. [17] Giuseppe Siracusano, Davide Sanvito, Roberto Gonzalez, Manikantan Srinivasan, Sivakaman Kamatchi, Wataru Takahashi, Masaru Kawakita, Takahiro Kakumaru, and Roberto Bifulco. 2023. Time for action: Automated analysis of cyber threat intelligence in the wild. arXiv preprint arXiv:2307.10214 (2023). [18] STIX Project. 2021. STIX. https://stixproject.github.io/. Accessed: 2026-05-15. [19] Mahzabin Tamanna, Shaswata Mitra, Md Erfan, Ahmed Ryan, Sudip Mittal, Laurie Williams, and Md Rayhanur Rahman. 2026. What Are Adversaries Doing? Automating Tactics, Techniques, and Procedures Extraction: A Systematic Review. arXiv:2604.02377 (2026). [20] Ming Xu, Hongtai Wang, Jiahao Liu, Xinfeng Li, Zhengmin Yu, Weili Han, Hoon Wei Lim, Jin Song Dong, and Jiaheng Zhang. 2024. ThreatPilot: Attack-Driven Threat Intelligence Extraction. arXiv:2412.10872 (2024).