XekRung Technical Report

arXiv:2605.00072 · cs.CR, cs.AI · Submitted 2026-04-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "XekRung Technical Report".

Elias: Here's the gist of the paper. XekRung is a frontier large language model for cybersecurity designed to provide comprehensive security capabilities by developing diverse data synthesis pipelines and a complete training pipeline spanning continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL) <ref:2605.00072#pg2>.

Nadia: Model Performance and Benchmarking XekRung achieves state-of-the-art performance on cybersecurity benchmarks among models of the same scale, while maintaining strong performance on general benchmarks <ref:2605.00072#pg2>.

Elias: The overall average score for XekRung-8B across 15 cybersecurity benchmarks is 81.04% <ref:2605.00072#pg2>.

Nadia: This represents a +9.03 percentage point improvement over its base model Qwen3-8B (72.01%) and surpasses all same-scale baselines by a substantial margin <ref:2605.00072#pg2>.

Title and authors: Nadia: Let's talk about the title and who wrote this, the XekRung Technical Report. It sets up what we are looking at—a large language model built for cybersecurity that aims to give it comprehensive security capabilities.

Elias: The authors are Jiutian Zeng, Junjie Li, Chengwei Dai, Jie Liang, Zhaoyu Hu, Yiliang Zhang, Ziang Weng, and Longtao Huang as core contributors. They seem to have a strong mix of expertise here.

Priya: I noticed the team includes people focused on different parts of the problem—from measurement research to exploitation and defense. That suggests they're trying to cover the whole cycle.

Nadia: Right, because what they're doing is tackling something inherently cross-domain and cross-layer, which means you need mastery over software engineering, operating systems, network protocols, and cryptographic primitives all at once.

Elias: The paper also points out that the search space in cybersecurity is vast because adversarial behaviors can show up across an enormous landscape of possible configurations and execution paths that just makes exhaustive enumeration infeasible.

Priya: That sounds like a huge hurdle for any system trying to learn this, so I'm curious if their data strategies really help navigate that complexity.

The paper's summary: Nadia: Now let's look at what the report actually summarizes about XekRung. Basically, they show that XekRung achieves state-of-the-art performance on cybersecurity benchmarks compared to other models of the same scale, while still keeping strong performance on general tasks.

Elias: The key number here is that the average score across fifteen cybersecurity benchmarks for XekRung-8B is eighty-one point zero four percent, which they say is a nine point zero three percentage point improvement over its base model, Qwen3-8B, which scored seventy-two point zero one percent.

Priya: eighty-one point zero four percent on those specific security metrics sounds pretty solid for a specialized model, but I need to know what that means in terms of real utility versus just high scores on a test set.

Nadia: It means it outperforms all the same-scale baselines by a substantial margin, and critically, XekRung-8B still holds seventy-two point five four percent overall on general purpose benchmarks, which is close to the base Qwen3-8B's seventy-three point thirty one percent.

Elias: That comparison shows that they managed to get deep domain expertise without completely losing general competence, which is a tough balance to strike in these specialized training efforts.

The paper's improvements: Nadia: The report highlights the specific ways they improved the model's capabilities. They built a system spanning continued pre-training, supervised fine-tuning, and reinforcement learning to extend the model’s abilities across tasks.

Elias: They introduce a multi-dimensional evaluation system that guides iterative improvement for both domain-specific and general purpose abilities throughout the whole training process.

Priya: I'm interested in how they structured that RL part because traditional RL often targets isolated subtasks, but this one seems to aim for something more unified.

Nadia: They move away from targeting just isolated subtasks by introducing a framework that unifies reinforcement learning with verifiable rewards, which we call Verifiable Rewards or RLVR for those specific tasks.

Elias: Plus, they use Agentic RL for end-to-end workflow optimization and they integrate reasoning reinforcement learning with preference alignment within a shared optimization loop.

Priya: That sounds like it’s trying to optimize the entire chain of security analysis, not just one step in it. How does that translate to actual capability?

Conclusion: Nadia: To wrap things up on this XekRung Technical Report, the main implication is that a well-trained 8B model can match or even exceed models with several times more parameters when it comes to domain-specific security tasks.

Elias: That highlights how effective domain compression can be, showing that deep expertise in a smaller model can outperform much larger models on specialized problems.

Priya: I just want to stress that the authors mention they are planning to further expand the model’s agentic capabilities and look into self-evolution for cybersecurity applications next.

Nadia: Right, so while this report shows strong performance, the team is focused on making it more capable of actually *doing* things autonomously in a security context.

Elias: They also plan to pursue broader cybersecurity research and applications based on this work, continuing the development of AI for Security.

Alibaba Security AGI Lab

cs.CR, cs.AI

Submitted: 2026-04-30

Updated: 2026-11-06

Comments: 23 pages, 2 figures, 5 tables. Jiutian Zeng, Junjie Li, Chengwei Dai, Jie Liang, and Zhaoyu Hu contributed equally to this work

Code: https://github.com/XuanwuAI/SecEval

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 100/100

The gist: XekRung is a frontier large language model for cybersecurity designed to provide comprehensive security capabilities by developing diverse data synthesis pipelines and a complete training pipeline…

Key concepts

Full Training Pipeline (CPT→SFT→RL)
This is a complete, three-stage training process: Continued Pre-Training (CPT) builds foundational knowledge on cybersecurity concepts; Supervised Fine-Tuning (SFT) teaches the model specific security tasks; and Reinforcement Learning (RL) refines behavior through trial and error. This ensures the model learns both general language skills and deep, specialized security expertise.
Group Relative Policy Optimization (GRPO)
GRPO is a specific type of multi-task reinforcement learning used during the training phase. It allows the model to optimize its performance across many different security tasks simultaneously by comparing its actions against a group of related policies. This helps the model learn robust strategies for various security scenarios efficiently.
Agentic RL and ActionRL
These concepts enable the model to act like an autonomous agent in complex workflows. Agentic RL optimizes end-to-end operational tasks, while ActionRL specifically focuses supervision on critical decision points, known as the 'Branch Divergence Point,' ensuring the model makes precise, effective moves during security analysis.

Terminology

Summary

XekRung is a frontier large language model for cybersecurity designed to provide comprehensive security capabilities by developing diverse data synthesis pipelines and a complete training pipeline spanning continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL) <ref:2605. p. 0

Where the paper says it

To achieve this, we develop diverse data synthesis pipelines tailored to the cybersecurity domain, enabling the scalable construction of high-quality training data and providing a strong foundation for cybersecurity knowledge and understanding. Building on this foundation, we establish a complete training pipeline spanning continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL) to further extend the model’s capabilities.

00072#pg2>.

Model Performance and Benchmarking

XekRung achieves state-of-the-art performance on cybersecurity benchmarks among models of the same scale, while maintaining strong performance on general benchmarks <ref:2605.00072#pg2>. The overall average score for XekRung-8B across 15 cybersecurity benchmarks is 81. p. 0

Where the paper says it

Extensive experiments demonstrate that XekRung achieves state-of-the-art performance on cybersecurity-specific benchmarks among models of the same scale, while maintaining strong performance on general benchmarks. Overall Average Score 15 Benchmarks Average across cybersecurity benchmarks CSEBench · CTIBench-MCQ/RCM/VSP/ATE · MMLU-CompSec · SecEval · SecBench-MCQ · CyberMetric10kQ/2kQ/500Q/80Q · CISSP-MC-ZH/SC-EN/SC-ZH Qwen3-8B Qwen3-14B Qwen3.

04% <ref:2605. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

00072#pg2>. This represents a +9.03 percentage point improvement over its base model Qwen3-8B (72. p. 2
Where the paper says it

04% across all 15 security benchmarks, representing a +9. 03 percentage point improvement over its base model Qwen3-8B (72.

01%) and surpasses all same-scale baselines by a substantial margin <ref:2605. p. 2
Where the paper says it

03 percentage point improvement over its base model Qwen3-8B (72. 01%) and surpassing all same-scale baselines by a substantial margin—including Foundation-Sec-8B-Reasoning (68.

00072#pg2>. Crucially, XekRung-8B retains 72. p. 2
Where the paper says it

21%) by over 12 percentage points on domain tasks. Crucially, these domain gains come with minimal degradation of general capabilities: XekRung-8B retains 72.

54% overall on general-purpose benchmarks, closely tracking the base Qwen3-8B (73. p. 2
Where the paper says it

Crucially, these domain gains come with minimal degradation of general capabilities: XekRung-8B retains 72. 54% overall on general-purpose benchmarks, closely tracking the base Qwen3-8B (73.

31%) <ref:2605. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

00072#pg2>.

Data Engineering and Synthesis Strategies

The authors built a data engineering system spanning the entire training lifecycle, combining learned quality scoring, capability-level interference balancing, and vector-space deduplication on the quality side <ref:2605.00072#pg2>. They introduced four targeted strategies to augment the corpus along the dimensions of representation, correlation, long-range dependency and engineering experience injection <ref:2605. p. 5

Where the paper says it

3 Domain-Specific Data Strategies While the unified pipeline resolves generic data quality issues, cybersecurity corpora present several domain-specific challenges: critical artifact types suffer from severe data sparsity; semantically related threat intelligence is fragmented across disparate sources; and genuine long-range dependencies are rarely captured by naive document concatenation. To address these, we introduce four targeted strategies to augment the corpus along the dimensions of representation, correlation, long-range dependency and engineering experience injection.

00072#pg2>. These strategies include:

  1. Explanatory Annotation for Domain-Specific Artifacts <ref:2605. p. 5

    Where the paper says it

    To address these, we introduce four targeted strategies to augment the corpus along the dimensions of representation, correlation, long-range dependency and engineering experience injection. Explanatory Annotation for Domain-Specific Artifacts.

    00072#pg7>.

  2. Correlated Information Aggregation. p. 5

    Where the paper says it

    To overcome this, we pair each raw artifact with a concise, synthetically generated natural-language explanation that articulates its semantics, security implications, and operational context, effectively transforming specialized tool outputs into self-contained, semantically rich training instances. Correlated Information Aggregation.

  3. Long-Context Data Synthesis via Entropy. p. 6

    Where the paper says it

    , 2023), and the two can be combined. Long-Context Data Synthesis via Entropy.

  4. Explicit Orchestration of Engineering Experience. p. 6

    Where the paper says it

    Crucially, the fundamental long-range capabilities established during this phase lay the structural foundation within the attention mechanism for processing genuine, long-horizon agent trajectories during mid-training (see § 3. Explicit Orchestration of Engineering Experience.

Complete and Innovative Training Pipeline

XekRung realizes a full CPT → SFT → RL pipeline with task-aware specialization for cybersecurity. The post-training unifies Reinforcement Learning with Verifiable Rewards (RLVR) for verifiable-reward tasks and Agentic RL for end-to-end workflow optimization. p. 2

Where the paper says it

XekRung realizes a full CPT → SFT → RL pipeline with task-aware specialization for cybersecurity. Post-training unifies Reinforcement Learning with Verifiable Rewards (RLVR) for verifiable-reward tasks and Agentic RL for end-to-end workflow optimization, complemented by code-verified adversarial self-evolution and vulnerability self-play.

This framework integrates Reasoning RL, Preference Alignment, and Agentic RL within a shared optimization loop. p. 2
Where the paper says it

, 2025) has demonstrated the potency of reinforcement learning for complex problem-solving, the systematic application of RL across the full breadth of cybersecurity tasks—with their heterogeneous reward structures, diverse task formulations, and stringent operational constraints—remains largely underexplored. Existing efforts, where RL is employed at all, tend to target isolated subtasks rather than pursuing a unified, multi-task RL framework capable of jointly optimizing analytical reasoning, agentic workflows, and behavioral alignment within the security domain.

Furthermore, the terminal mechanism relies upon sustained, autonomous self-improvement through Self-Evolving via Adversarial RL with Code-Grounded Verification. p. 8
Where the paper says it

Building upon this foundation, it advances to a multi-task RL phase powered by Group Relative Policy Optimization (GRPO). Ultimately, a self-evolutionary loop grounded in verifiable code execution ensures sustained breakthroughs in complex reasoning.

Evaluation Protocol and Results

The comprehensive evaluation involves a multi-dimensional protocol covering 35 general and 15 cybersecurity-specific benchmarks. The results show that XekRung-8B outperforms both general-purpose baselines and security-specialized baselines across both benchmark suites. p. 0

Where the paper says it

04 ≈57 Figure 1: Overview of XekRung’s performance across 15 cybersecurity benchmarks compared with general-purpose frontier models and existing security-specialized models. XekRung achieves leading results on all evaluated dimensions.

Specifically, at the 8B parameter scale, XekRung-8B outperforms both general-purpose baselines (Qwen3-8B, Llama-3. p. 2
Where the paper says it

03 percentage point improvement over its base model Qwen3-8B (72. 01%) and surpassing all same-scale baselines by a substantial margin—including Foundation-Sec-8B-Reasoning (68.

1-8B) and security-specialized baselines (Foundation-Sec-8BReasoning, Llama-PrimusReasoning) across both benchmark suites. p. 0
Where the paper says it

3-70B-Instruct Llama-Primus-Reasoning-8B Foundation-Sec-8B-Reasoning SecGPT-14B XekRung-8B 72. 04 ≈57 Figure 1: Overview of XekRung’s performance across 15 cybersecurity benchmarks compared with general-purpose frontier models and existing security-specialized models.

Scaling Implications

XekRung achieves competitive or superior performance relative to Qwen3-14B (nearly 2× the parameter count) across several benchmark categories. XekRung-8B also outperforms the larger SecGPT-14B on security benchmarks by a significant margin (81. p. 0

Where the paper says it

3-70B-Instruct Llama-Primus-Reasoning-8B Foundation-Sec-8B-Reasoning SecGPT-14B XekRung-8B 72. 04 ≈57 Figure 1: Overview of XekRung’s performance across 15 cybersecurity benchmarks compared with general-purpose frontier models and existing security-specialized models.

04 vs. 68.21 overall). p. 0
Where the paper says it

Extensive experiments demonstrate that XekRung achieves state-of-the-art performance on cybersecurity-specific benchmarks among models of the same scale, while maintaining strong performance on general benchmarks. Overall Average Score 15 Benchmarks Average across cybersecurity benchmarks CSEBench · CTIBench-MCQ/RCM/VSP/ATE · MMLU-CompSec · SecEval · SecBench-MCQ · CyberMetric10kQ/2kQ/500Q/80Q · CISSP-MC-ZH/SC-EN/SC-ZH Qwen3-8B Qwen3-14B Qwen3.

This highlights the high cost-effectiveness of domain compression, as a well-trained 8B model can match or exceed models with several times more parameters on domain-specific tasks. p. 2
Where the paper says it

We systematize evaluation into a multi-dimensional protocol covering 35 general and 15 cybersecurity-specific benchmarks, jointly assessing domain proficiency and generalcapability retention to resist narrow-task overfitting. XekRung-8B achieves state-of-the-art security performance among same-scale models while competing with several-times-larger models in nonthinking mode with exceptional token efficiency.

Each training stage contributes meaningfully to the final model, ensuring deep domain expertise while maintaining broad general-purpose competence. p. 6
Where the paper says it

The final mixture is selected by jointly considering domain performance, general-capability retention, and training stability rather than any single metric. Overall, we follow the mixing principle of preserving general competence while strengthening cybersecurity expertise, enabling the model to substantially absorb domain knowledge without materially sacrificing its general understanding and reasoning ability.

Conclusion

In this technical report, we introduce XekRung, a frontier large language model tailored for the cybersecurity. The authors plan to further expand the model’s agentic capabilities and investigate the potential of self-evolution for cybersecurity. p. 2

Where the paper says it

To bridge the gap between surface-level knowledge and deep operational understanding (the core limitation of general-purpose models identified above), we devise domain-aware data construction and synthesis strategies that enrich rare yet decision-critical security concepts, model cross-artifact dependencies, and transform opaque security artifacts into self-explanatory training instances— collectively enabling the model to internalize the structural and causal patterns of the domain rather than merely memorizing isolated facts. To overcome the limitations of narrow task-specific RL (the key deficit of existing domain models), we design a unified multi-task RL framework that integrates Reasoning RL with verifiable rewards for analytical tasks and Agentic RL for multi-step tool-augmented operational workflows, complemented by an adversarial self-evolution mechanism for continuous capability escalation.

They also plan to pursue broader cybersecurity research and applications, contributing to the continued advancement of AI for Security.

Authors

Core Contributors are Jiutian Zeng, Junjie Li, Chengwei Dai, Jie Liang, Zhaoyu Hu, Yiliang Zhang, Ziang Weng, Longtao Huang. Other contributors include Dongjie Zhang and Libin Dong. p. 3

Where the paper says it

For comprehensive overviews of LLMs in cybersecurity, we refer readers to recent surveys (Xu et al. , 2025; Zhang et al.

References

Rishabh Agarwal et al. On-policy distillation of language models: Learning from self-generated mistakes. p. 5

Where the paper says it

In their raw form, these artifacts manifest as domain-specific sequences of memory addresses, hexadecimal bytes, or domain-specific language (DSL) tokens, rendering direct representation learning highly inefficient. To overcome this, we pair each raw artifact with a concise, synthetically generated natural-language explanation that articulates its semantics, security implications, and operational context, effectively transforming specialized tool outputs into self-contained, semantically rich training instances.

In The Twelfth International Conference on Learning Representations, 2024. p. 3
Where the paper says it

, 2025) introduced Group Relative Policy Optimization (GRPO) (Shao et al. , 2024), which replaces a learned critic with intra-group relative rewards and allows deliberate reasoning behavior to emerge organically during reinforcement learning.

Md Tanvirul Alam et al. CTIBench: A benchmark for evaluating LLMs in cyber threat intelligence. p. 2
Where the paper says it

To accommodate heterogeneous computational budgets, we additionally devise a knowledge distillation pipeline that transfers aligned capabilities into compact variants (≤4B). As depicted in Figure 1, we carry out extensive evaluation on a multi-dimensional benchmark suite comprising 35 general-purpose benchmarks across 9 categories and 15 cybersecurity-specific benchmarks spanning threat intelligence, vulnerability analysis, security knowledge, and certification.

In The Thirty-eighth Conference on Neural Information Processing Systems, Datasets and Benchmarks Track, 2024. Anthropic. Assessing Claude Mythos Preview’s cybersecurity capabilities, 2026 URL https://red. p. 1
Where the paper says it

This depth of domain comprehension—the capacity to fluidly integrate cross-layer knowledge, perform multi-step causal inference, and translate analytical conclusions into concrete security actions—cannot be reliably elicited through prompting or context augmentation alone; it necessitates the model having internalized the structural and causal patterns of the domain through dedicated training. Most recently, frontier proprietary models have begun to turn their attention squarely to cybersecurity: Anthropic’s Claude Mythos Preview (Anthropic, 2026) exhibits emergent offensive capabilities such as zero-day discovery and autonomous exploit chain construction, while OpenAI’s GPT-5.

anthropic.com/2026/mythos-preview/. p. 1
Where the paper says it

This depth of domain comprehension—the capacity to fluidly integrate cross-layer knowledge, perform multi-step causal inference, and translate analytical conclusions into concrete security actions—cannot be reliably elicited through prompting or context augmentation alone; it necessitates the model having internalized the structural and causal patterns of the domain through dedicated training. Most recently, frontier proprietary models have begun to turn their attention squarely to cybersecurity: Anthropic’s Claude Mythos Preview (Anthropic, 2026) exhibits emergent offensive capabilities such as zero-day discovery and autonomous exploit chain construction, while OpenAI’s GPT-5.

Jinze Bai et al. p. 3
Where the paper says it

1-8B-Base in particular demonstrating robust capabilities in reasoning and code generation that make it a natural initialization point for securityfocused models. Alibaba’s Qwen series (Bai et al.

Qwen technical report arXiv preprint arXiv:2309. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

16609, 2023 URL https://arxiv. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

org/abs/2309.16609. BaronSecurity. p. 3
Where the paper says it

, 2025), and Colibri (ColibriAI, 2025) show that task-specific supervision can yield strong performance on welldefined problems such as phishing analysis or vulnerability classification, though the relatively small data scales involved (typically 14K–55K samples) and the absence of deep knowledge injection often constrain generalization. More recent models have pushed further by appending reinforcement learning stages: BaronLLM (BaronSecurity, 2025), SecurityLLM (ZySec-AI , 2024), and Foundation-Sec-8BReasoning (Yang et al.

BaronLLM: Offensive security language model HuggingFace Model Repository, 2025 URL https://huggingface. p. 1
Where the paper says it

This depth of domain comprehension—the capacity to fluidly integrate cross-layer knowledge, perform multi-step causal inference, and translate analytical conclusions into concrete security actions—cannot be reliably elicited through prompting or context augmentation alone; it necessitates the model having internalized the structural and causal patterns of the domain through dedicated training. Most recently, frontier proprietary models have begun to turn their attention squarely to cybersecurity: Anthropic’s Claude Mythos Preview (Anthropic, 2026) exhibits emergent offensive capabilities such as zero-day discovery and autonomous exploit chain construction, while OpenAI’s GPT-5.

co/BaronSecurity/baron-llm. p. 3
Where the paper says it

, 2025), and Colibri (ColibriAI, 2025) show that task-specific supervision can yield strong performance on welldefined problems such as phishing analysis or vulnerability classification, though the relatively small data scales involved (typically 14K–55K samples) and the absence of deep knowledge injection often constrain generalization. More recent models have pushed further by appending reinforcement learning stages: BaronLLM (BaronSecurity, 2025), SecurityLLM (ZySec-AI , 2024), and Foundation-Sec-8BReasoning (Yang et al.

Yonatan Bisk et al. Piqa: Reasoning about physical commonsense in natural language Proceedings of AAAI, 2020. p. 5
Where the paper says it

In their raw form, these artifacts manifest as domain-specific sequences of memory addresses, hexadecimal bytes, or domain-specific language (DSL) tokens, rendering direct representation learning highly inefficient. To overcome this, we pair each raw artifact with a concise, synthetically generated natural-language explanation that articulates its semantics, security implications, and operational context, effectively transforming specialized tool outputs into self-contained, semantically rich training instances.

Cody Blakeney et al. Does your data spark joy? p. 0
Where the paper says it

May 4, 2026 XekRung Technical Report Alibaba Security AGI Lab Abstract We present XekRung, a frontier large language model for cybersecurity, designed to provide comprehensive security capabilities. To achieve this, we develop diverse data synthesis pipelines tailored to the cybersecurity domain, enabling the scalable construction of high-quality training data and providing a strong foundation for cybersecurity knowledge and understanding.

performance gains from domain upsampling at the end of training CoRR, abs/2406. p. 2
Where the paper says it

21%) by over 12 percentage points on domain tasks. Crucially, these domain gains come with minimal degradation of general capabilities: XekRung-8B retains 72.

03476, 2024 doi: 10. p. 1
Where the paper says it

Model capabilities are evolving beyond elementary dialogue and factual recall toward assisting humans in tackling concrete, high-complexity real-world problems (Grattafiori et al. , 2024; Qwen Team, 2025a; DeepSeek-AI, 2024; Jaech et al.

48550/ARXIV. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

2406.03476 URL https://doi. p. 5
Where the paper says it

, names, email addresses, phone numbers), anonymize internal identifiers and proprietary asset references, and selectively truncate or remove long, semantically vacuous encoded strings (e. , Base64 blocks, URL tracking parameters, session tokens).

org/10.48550/arXiv. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

2406.03476. Thomas Borsani et al. Gradient similarity surgery in multi-task deep learning In Proceedings of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECMLPKDD), 2025 arXiv:2506. p. 3
Where the paper says it

, 2025), and Colibri (ColibriAI, 2025) show that task-specific supervision can yield strong performance on welldefined problems such as phishing analysis or vulnerability classification, though the relatively small data scales involved (typically 14K–55K samples) and the absence of deep knowledge injection often constrain generalization. More recent models have pushed further by appending reinforcement learning stages: BaronLLM (BaronSecurity, 2025), SecurityLLM (ZySec-AI , 2024), and Foundation-Sec-8BReasoning (Yang et al.

06130. Lichang Chen et al. Alpagasus: Training a better alpaca with fewer data arXiv preprint arXiv:2307. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

08701, 2023a. Wenhu Chen et al. Theoremqa: A theorem-driven question answering dataset Proceedings of EMNLP, 2023b. p. 7
Where the paper says it

Combined with the progressive long-context schedule described above for agentic trajectories, repository-level samples further consolidate the model’s ability to maintain structural understanding across long sequences, laying the foundation for code-agent and security-audit capabilities acquired in post-training. 3 Linearized SFT Data Integration We integrate a diverse collection of SFT datasets—spanning cybersecurity question answering, code understanding, vulnerability analysis, and general instruction-following—by stripping chat templates and role markers, reformulating each sample as a plain-text continuation sequence, and training it under 8

Christopher Clark et al. Boolq: Exploring the surprising difficulty of natural yes/no questions Proceedings of NAACL, 2019. p. 1
Where the paper says it

, 2024; Qwen Team, 2025a; DeepSeek-AI, 2024; Jaech et al. As reasoning and code generation proficiencies continue to mature, identifying application domains of sufficient difficulty has itself become a central proposition for probing and extending the capability frontier of LLMs.

Peter Clark et al. Think you have solved question answering? p. 7
Where the paper says it

Combined with the progressive long-context schedule described above for agentic trajectories, repository-level samples further consolidate the model’s ability to maintain structural understanding across long sequences, laying the foundation for code-agent and security-audit capabilities acquired in post-training. 3 Linearized SFT Data Integration We integrate a diverse collection of SFT datasets—spanning cybersecurity question answering, code understanding, vulnerability analysis, and general instruction-following—by stripping chat templates and role markers, reformulating each sample as a plain-text continuation sequence, and training it under 8

try arc, the ai2 reasoning challenge arXiv preprint arXiv:1803. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

05457, 2018. Clouditera. p. 3
Where the paper says it

, 2026) each combine SFT with RL to improve alignment and reasoning reliability, though none begins from a domain-pretrained checkpoint, leaving a gap in foundational domain coverage. 0 (Clouditera, 2025) is the closest predecessor to our own pipeline, employing CPT followed by SFT and RL on Qwen2.

Secgpt: The world’s first open-source large language model for cybersecurity https://huggingface. p. 4
Where the paper says it

We organize the data pipeline into four stages: collection, processing, domain-specific data strategies, and data mixture. 1 Data Collection We assemble a large-scale, multi-source cybersecurity corpus that spans five complementary categories: • Open-source and external data.

co/clouditera/secgpt, 2025 Accessed: 2025-04. p. 4
Where the paper says it

1-8B SFT → RL SecGPT-2. 0 (Clouditera, 2025) Qwen2.

Qoder Team. p. 1
Where the paper says it

Model capabilities are evolving beyond elementary dialogue and factual recall toward assisting humans in tackling concrete, high-complexity real-world problems (Grattafiori et al. , 2024; Qwen Team, 2025a; DeepSeek-AI, 2024; Jaech et al.

The next evolution toward intelligent editing: Qoder NEXT model and ActionRL preference alignment in practice, 2026 URL https://qoder. p. 6
Where the paper says it

This enables the model to internalize the endto-end reasoning chain—spanning fault discovery, root-cause localization, and patch generation— that epitomizes expert software engineering practice. This approach aligns with recent findings in reinforcement learning from software evolution (Wei et al.

com/blog/qoder-next-model? p. 0
Where the paper says it

May 4, 2026 XekRung Technical Report Alibaba Security AGI Lab Abstract We present XekRung, a frontier large language model for cybersecurity, designed to provide comprehensive security capabilities. To achieve this, we develop diverse data synthesis pipelines tailored to the cybersecurity domain, enabling the scalable construction of high-quality training data and providing a strong foundation for cybersecurity knowledge and understanding.

type=%E6%8A%80%E6%9C%AF Accessed: 2025. p. 1
Where the paper says it

Second, the associated search space is vast: adversarial behaviors can manifest across an enormous landscape of possible configurations, inputs, and execution paths, rendering exhaustive enumeration infeasible and demanding strategic, heuristic-guided exploration. Third, and most critically, causal chains are often deeply concealed and counterintuitive—a security vulnerability may stem from subtle interactions, such as a type confusion triggered under a rare compiler optimization path or a privilege escalation enabled by an undocumented kernel behavior, that demand multi-step reasoning to identify and resolve.

Qwen Team. Qwen3 technical report arXiv preprint arXiv:2505. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

09388, 2025a URL https://arxiv. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

org/abs/2505.09388. Qwen Team. Qwen3.5 technical report arXiv preprint arXiv:2509. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

12345, 2025b URL https://arxiv. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

org/abs/2509.12345. Rafael Rafailov et al. Direct preference optimization: Your language model is secretly a reward model In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine (eds.), Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023 NeurIPS 2023, New Orleans, LA, USA December 10 - 16, 2023 URL http://papers.nips.cc/paper/paper/a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference. p. 0
Where the paper says it

May 4, 2026 XekRung Technical Report Alibaba Security AGI Lab Abstract We present XekRung, a frontier large language model for cybersecurity, designed to provide comprehensive security capabilities. To achieve this, we develop diverse data synthesis pipelines tailored to the cybersecurity domain, enabling the scalable construction of high-quality training data and providing a strong foundation for cybersecurity knowledge and understanding.

html. Pranav Rajpurkar et al. Know what you don’t know: Unanswerable questions for squad Proceedings of ACL, 2018. David Rein et al. Gpqa: A graduate-level google-proof q&a benchmark Proceedings of ICLR, 2024. p. 2
Where the paper says it

To accommodate heterogeneous computational budgets, we additionally devise a knowledge distillation pipeline that transfers aligned capabilities into compact variants (≤4B). As depicted in Figure 1, we carry out extensive evaluation on a multi-dimensional benchmark suite comprising 35 general-purpose benchmarks across 9 categories and 15 cybersecurity-specific benchmarks spanning threat intelligence, vulnerability analysis, security knowledge, and certification.

Keisuke Sakaguchi et al. Winogrande: An adversarial winograd schema challenge at scale Proceedings of AAAI, 2021. p. 8
Where the paper says it

4 Post-Training This section provides a comprehensive account of the post-training technical framework. The pipeline systematically integrates rigorous data quality control, scalable synthetic data generation, and advanced alignment methodologies, encompassing supervised fine-tuning (SFT), multi-task reinforcement learning (RL), and adversarial self-evolution mechanisms, as well as structured knowledge distillation that transfers the aligned capabilities of large-scale models into compact models for efficient deployment.

Maarten Sap et al. Social iqa: Commonsense reasoning about social interactions Proceedings of EMNLP, 2019. p. 1
Where the paper says it

Second, the associated search space is vast: adversarial behaviors can manifest across an enormous landscape of possible configurations, inputs, and execution paths, rendering exhaustive enumeration infeasible and demanding strategic, heuristic-guided exploration. Third, and most critically, causal chains are often deeply concealed and counterintuitive—a security vulnerability may stem from subtle interactions, such as a type confusion triggered under a rare compiler optimization path or a privilege escalation enabled by an undocumented kernel behavior, that demand multi-step reasoning to identify and resolve.

Tom Schaul et al. Prioritized experience replay In Yoshua Bengio and Yann LeCun (eds. p. 4
Where the paper says it

The corpus consists of heterogeneous cybersecurity data from multiple sources, curated through a unified pipeline of cleaning, augmentation, and format normalization. On top of the curated corpus, we introduce code-centric signals that explicitly encode engineering experience, thereby improving the efficiency with which tacit domain expertise is absorbed.

), 4th International Conference on Learning Representations, ICLR 2016 San Juan, Puerto Rico May 2-4, 2016 Conference Track Proceedings, 2016 URL http://arxiv.org/abs/1511.05952. Muxi Diao et al. SEAS: self-evolving adversarial safety optimization for large language models In Toby Walsh, Julie Shah, and Zico Kolter (eds. p. 1
Where the paper says it

1 Introduction The development of large language models (LLMs) has entered a pivotal phase. Model capabilities are evolving beyond elementary dialogue and factual recall toward assisting humans in tackling concrete, high-complexity real-world problems (Grattafiori et al.

), Thirty-Ninth AAAI Conference on Artificial Intelligence, Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence, Fifteenth Symposium on Educational Advances in Artificial Intelligence AAAI 2025 Philadelphia, PA USA February 25 - March 4, 2025 pp. 23778–23786 AAAI Press, 2025 doi: 10. p. 1
Where the paper says it

Foundation-Sec8B (Kassianik et al. , 2025) conducts continued pre-training atop Llama; Primus (Yu et al.

1609/AAAI.V39I-N. Alexandra Dragomir et al. Clewr: Curriculum learning with restarts for machine translation preference learning CoRR, abs/2601. p. 7
Where the paper says it

Following a breadth-first, specialization-second curriculum principle, we therefore reserve long-horizon agentic data for injection during mid-training. Two-Dimensional Curriculum Learning.

05858, 2026 doi: 10. p. 0
Where the paper says it

May 4, 2026 XekRung Technical Report Alibaba Security AGI Lab Abstract We present XekRung, a frontier large language model for cybersecurity, designed to provide comprehensive security capabilities. To achieve this, we develop diverse data synthesis pipelines tailored to the cybersecurity domain, enabling the scalable construction of high-quality training data and providing a strong foundation for cybersecurity knowledge and understanding.

48550/ARXIV.2601.05858 URL https://doi. p. 5
Where the paper says it

, names, email addresses, phone numbers), anonymize internal identifiers and proprietary asset references, and selectively truncate or remove long, semantically vacuous encoded strings (e. , Base64 blocks, URL tracking parameters, session tokens).

org/10.48550/arXiv.2601.05858. Dheeru Dua et al. Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs Proceedings of NAACL, 2019. p. 1
Where the paper says it

Security tasks frequently require not just recalling that a particular vulnerability class exists, but understanding the precise conditions under which it manifests, reasoning about its exploitability across interacting system components, and synthesizing actionable remediation strategies grounded in operational constraints. This depth of domain comprehension—the capacity to fluidly integrate cross-layer knowledge, perform multi-step causal inference, and translate analytical conclusions into concrete security actions—cannot be reliably elicited through prompting or context augmentation alone; it necessitates the model having internalized the structural and causal patterns of the domain through dedicated training.

Chaochen Gao et al. Nextlong: Toward effective longcontext training without long documents In Proceedings of the Forty-Second International Conference on Machine Learning (ICML), 2025. p. 7
Where the paper says it

Following a breadth-first, specialization-second curriculum principle, we therefore reserve long-horizon agentic data for injection during mid-training. Two-Dimensional Curriculum Learning.

Saurav Garg et al. Mmluredux: A closer look at mmlu arXiv preprint arXiv:2406. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

02607, 2024. p. 1
Where the paper says it

Model capabilities are evolving beyond elementary dialogue and factual recall toward assisting humans in tackling concrete, high-complexity real-world problems (Grattafiori et al. , 2024; Qwen Team, 2025a; DeepSeek-AI, 2024; Jaech et al.

Aaron Grattafiori et al. p. 1
Where the paper says it

1 Introduction The development of large language models (LLMs) has entered a pivotal phase. Model capabilities are evolving beyond elementary dialogue and factual recall toward assisting humans in tackling concrete, high-complexity real-world problems (Grattafiori et al.

The Llama 3 herd of models arXiv preprint arXiv:2407. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

21783, 2024 URL https://arxiv. p. 0
Where the paper says it

XekRung achieves leading results on all evaluated dimensions. 1 arXiv:2605.

org/abs/2407.21783. Yizhong Wang et al. p. 9
Where the paper says it

This compels the student to actively diagnose and correct its native generation patterns, yielding robust improvements in out-of-distribution generalization. For the explicit distillation objective, we supplant traditional KL divergence with ABKD (Wang et al.

Self-instruct: Aligning language models with self-generated instructions In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), 2023. p. 5
Where the paper says it

In their raw form, these artifacts manifest as domain-specific sequences of memory addresses, hexadecimal bytes, or domain-specific language (DSL) tokens, rendering direct representation learning highly inefficient. To overcome this, we pair each raw artifact with a concise, synthetically generated natural-language explanation that articulates its semantics, security implications, and operational context, effectively transforming specialized tool outputs into self-contained, semantically rich training instances.

Yubo Wang et al. p. 9
Where the paper says it

This compels the student to actively diagnose and correct its native generation patterns, yielding robust improvements in out-of-distribution generalization. For the explicit distillation objective, we supplant traditional KL divergence with ABKD (Wang et al.

Mmlu-pro: A more robust and challenging multi-task language understanding benchmark arXiv preprint arXiv:2406.01564, 2024. p. 1
Where the paper says it

Model capabilities are evolving beyond elementary dialogue and factual recall toward assisting humans in tackling concrete, high-complexity real-world problems (Grattafiori et al. , 2024; Qwen Team, 2025a; DeepSeek-AI, 2024; Jaech et al.

How the training signal is built

The authors devise a complete three-stage training pipeline spanning Continued Pre-Training (CPT), Supervised Fine-Tuning (SFT), and Multi-Task Reinforcement Learning. The pre-training stage exposes the base model to a curated cybersecurity corpus, establishing foundational understanding of security concepts, terminology, and ontology. p. 4

Where the paper says it

Our training curriculum further separates base pre-training from a dedicated mid-training stage, in which higher-order capabilities—long-horizon agentic planning, supervised task knowledge, and repository-level long-context code understanding—are injected without compromising the general competence established in the base stage. 1 Data Curation and Processing Building a cybersecurity-specialized foundation model requires a training corpus that captures both the breadth and depth of the domain, ranging from high-level threat intelligence narratives to lowlevel binary analysis artifacts.

The mid-training stage injects higher-order capabilities without compromising general competence. p. 4
Where the paper says it

On top of the curated corpus, we introduce code-centric signals that explicitly encode engineering experience, thereby improving the efficiency with which tacit domain expertise is absorbed. Our training curriculum further separates base pre-training from a dedicated mid-training stage, in which higher-order capabilities—long-horizon agentic planning, supervised task knowledge, and repository-level long-context code understanding—are injected without compromising the general competence established in the base stage.

The post-training transitions rigorously from behavioral cloning to reinforced exploration through SFT, multi-task RL powered by Group Relative Policy Optimization (GRPO), and a self-evolutionary loop grounded in verifiable code execution. p. 8
Where the paper says it

Building upon this foundation, it advances to a multi-task RL phase powered by Group Relative Policy Optimization (GRPO). Ultimately, a self-evolutionary loop grounded in verifiable code execution ensures sustained breakthroughs in complex reasoning.

Where the gains come from

The gains stem from the synergy of a well-integrated, full-pipeline training paradigm, domain-tailored data engineering, and multi-task RL. The methodology is predicated on extreme capability decomposition and compositional generalization. p. 2

Where the paper says it

XekRung undergoes a complete training pipeline spanning Continued Pre-Training (CPT), Supervised Fine-Tuning (SFT), and Multi-Task Reinforcement Learning, with each stage engineered around the unique data characteristics and task diversity of the cybersecurity domain. Our methodology is predicated on extreme capability decomposition and compositional generalization: we systematically disentangle the core task dimensions of cybersecurity— vulnerability analysis, threat intelligence, secure code engineering, and agentic security operations—and reassemble them through targeted training that engenders mutual reinforcement across dimensions.

They devise domain-aware data construction and synthesis strategies that enrich rare yet decision-critical security concepts, model cross-artifact dependencies, and transform opaque security artifacts into self-explanatory training instances. p. 2
Where the paper says it

Our methodology is predicated on extreme capability decomposition and compositional generalization: we systematically disentangle the core task dimensions of cybersecurity— vulnerability analysis, threat intelligence, secure code engineering, and agentic security operations—and reassemble them through targeted training that engenders mutual reinforcement across dimensions. To bridge the gap between surface-level knowledge and deep operational understanding (the core limitation of general-purpose models identified above), we devise domain-aware data construction and synthesis strategies that enrich rare yet decision-critical security concepts, model cross-artifact dependencies, and transform opaque security artifacts into self-explanatory training instances— collectively enabling the model to internalize the structural and causal patterns of the domain rather than merely memorizing isolated facts.

The multi-task RL framework integrates Reasoning RL with Verifiable Rewards (RLVR) for verifiable-reward tasks and Agentic RL for multi-step tool-augmented operational workflows. p. 2
Where the paper says it

To bridge the gap between surface-level knowledge and deep operational understanding (the core limitation of general-purpose models identified above), we devise domain-aware data construction and synthesis strategies that enrich rare yet decision-critical security concepts, model cross-artifact dependencies, and transform opaque security artifacts into self-explanatory training instances— collectively enabling the model to internalize the structural and causal patterns of the domain rather than merely memorizing isolated facts. To overcome the limitations of narrow task-specific RL (the key deficit of existing domain models), we design a unified multi-task RL framework that integrates Reasoning RL with verifiable rewards for analytical tasks and Agentic RL for multi-step tool-augmented operational workflows, complemented by an adversarial self-evolution mechanism for continuous capability escalation.

Furthermore, the self-play framework orchestrates an iterative engagement between a Red Team model (attacker) and a Target model (defender). p. 2
Where the paper says it

Existing efforts, where RL is employed at all, tend to target isolated subtasks rather than pursuing a unified, multi-task RL framework capable of jointly optimizing analytical reasoning, agentic workflows, and behavioral alignment within the security domain. To address these challenges, we introduce XekRung, a cybersecurity-specialized large language model built upon the Qwen foundation (Qwen Team, 2025a).

The continuous capability escalation is guaranteed by an exponentially expanding curriculum strictly defined by high-fidelity security permutations. p. 2
Where the paper says it

To overcome the limitations of narrow task-specific RL (the key deficit of existing domain models), we design a unified multi-task RL framework that integrates Reasoning RL with verifiable rewards for analytical tasks and Agentic RL for multi-step tool-augmented operational workflows, complemented by an adversarial self-evolution mechanism for continuous capability escalation. The current release, XekRung, is architected as a non-thinking model that directly emits high-fidelity responses without explicit intermediate reasoning traces, delivering stateof-the-art results with exceptional token efficiency—a decisive advantage for latency-sensitive security operations such as real-time alert triage and interactive threat analysis.

In this technical report, we introduce XekRung, a frontier large language model tailored for the cybersecurity. The authors will further expand the model’s agentic capabilities and investigate the potential of self-evolution for cybersecurity. p. 2

Where the paper says it

To bridge the gap between surface-level knowledge and deep operational understanding (the core limitation of general-purpose models identified above), we devise domain-aware data construction and synthesis strategies that enrich rare yet decision-critical security concepts, model cross-artifact dependencies, and transform opaque security artifacts into self-explanatory training instances— collectively enabling the model to internalize the structural and causal patterns of the domain rather than merely memorizing isolated facts. To overcome the limitations of narrow task-specific RL (the key deficit of existing domain models), we design a unified multi-task RL framework that integrates Reasoning RL with verifiable rewards for analytical tasks and Agentic RL for multi-step tool-augmented operational workflows, complemented by an adversarial self-evolution mechanism for continuous capability escalation.

They also plan to pursue broader cybersecurity research and applications, contributing to the continued advancement of AI for Security.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Data Synthesis Pipelines for Cybersecurity Knowledge

Improve AI systems by developing diverse data synthesis pipelines tailored to cybersecurity, enabling high-quality training data construction through contextual enrichment and targeted synthesis pipelines on the construction side. This allows models to internalize the structural and causal patterns of the domain rather than merely memorizing isolated facts.

  1. Complete Training Pipeline for Domain Specialization

Implement a full CPT to SFT to RL pipeline with task-aware specialization, specifically using Group Relative Policy Optimization (GRPO) for multi-task reinforcement learning. This allows the model to achieve deep domain expertise while maintaining broad general-purpose competence by optimizing Reasoning RL with Verifiable Rewards (RLVR).

  1. Enhanced Reasoning and Tool Orchestration

The system can perform complex, multi-step security analysis tasks such as automated penetration testing, large-scale vulnerability detection and validation, red-team long-chain attack simulation, and multi-stage incident response. This is achieved through Agentic RL for end-to-end workflow optimization complemented by ActionRL to focus supervision on the Branch Divergence Point.

  1. Cross-Scale Competitiveness in Deployment

The improved AI can achieve state of the art performance on cybersecurity benchmarks while maintaining strong general performance, demonstrating that a model can match or exceed models with several times more parameters on domain-specific tasks, while offering substantially faster inference—a critical advantage for latency-sensitive security operations.

  1. Robust Vulnerability Root Cause Mapping

The model can perform specific analytical tasks like Root Cause Mapping (RCM): Predicting CWE identifiers from CVE semantics, evaluated via a strict binary exact-match indicator and Vulnerability Severity Prediction (VSP): Forecasting aggregate CVSS scores alongside granular metric components.

Abstract

We present XekRung, a frontier large language model for cybersecurity, designed to provide comprehensive security capabilities. To achieve this, we develop diverse data synthesis pipelines tailored to the cybersecurity domain, enabling the scalable construction of high-quality training data and providing a strong foundation for cybersecurity knowledge and understanding. Building on this foundation, we establish a complete training pipeline spanning continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL) to further extend the model's capabilities. We further introduce a multi-dimensional evaluation system to guide the iterative improvement of both domain-specific and general-purpose abilities. Extensive experiments demonstrate that XekRung achieves state-of-the-art performance on cybersecurity-specific benchmarks among models of the same scale, while maintaining strong performance on general benchmarks.

Sources

Related papers