Cyber threat intelligence (CTI) reports contain rich information about advanced persistent threat (APT) groups. This information is useful for adversary profile construction. However, turning long and fragmented CTI reports into structured adversary profiles remains difficult. Existing methods mainly rely on named entity and relation extraction pipelines or single large language models (LLMs). These methods often lack explicit semantic constraints for profile-oriented tasks. They are also prone to omission and hallucination when processing long reports. In this paper, we study APT adversary profile construction from CTI reports. We formulate this task as an ontology-constrained profile information extraction task. To support this task, we develop VICTOR, a domain-specific ontology that organizes profile-relevant information into six dimensions. We then propose Debapt, an ontology-driven multi-agent debate framework. Guided by VICTOR, Debapt performs APT adversary profile construction through two debate loops in the entity extraction phase and relation extraction phase. In each loop, role-specialized agents iteratively extract, review, and adjudicate candidate profile-relevant facts under moderator supervision. These extracted results can be further aggregated across reports to support actor-centric adversary profile construction. To evaluate Debapt, we re-annotate three public CTI datasets under a unified profiling schema. Experimental results show that Debapt improves profile-oriented entity and relation extraction over competitive baselines. It extracts more complete and better grounded profile-relevant facts from CTI reports. Case studies further show that these extracted results can support the construction of analytically useful adversary profiles.

Cyber Threat Intelligence (CTI) parsing aims to extract key threat information from massive data, transform it into actionable intelligence, enhance threat detection and defense efficiency, including attack graph construction, intelligence fusion, and indicator extraction. Among these research topics, Attack Graph Construction (AGC) is essential for visualizing and understanding the potential attack paths of threat events from CTI reports. Existing approaches primarily construct the attack graphs purely from the textual data to reveal the logical threat relationships between entities within the attack behavioral sequence. However, they typically overlook the specific threat information inherent in visual modalities, which preserves key threat details from inherently multimodal CTI reports. Inspired by the remarkable multimodal understanding capabilities of Multimodal Large Language Models (MLLMs), we explore their potential in enhancing multimodal attack graph construction. To be specific, we propose a novel framework, MM-AttacKG, which can effectively extract key information from threat images and integrate it into attack graph construction, thereby enhancing the comprehensiveness and accuracy of attack graphs. It first employs a threat image parsing module to extract critical threat information from images and generate textual descriptions using MLLMs. Subsequently, it builds an iterative question-answering pipeline tailored for image parsing to refine the understanding of threat images. Finally, it achieves content-level integration between attack graphs and image-based answers through MLLMs, completing threat information enhancement. We construct a new multimodal dataset, AG-LLM-mm, and conduct extensive experiments to evaluate the effectiveness of MM-AttacKG. The results demonstrate that MM-AttacKG can accurately identify key information in threat images and significantly improve the quality of multimodal attack graph construction, effectively addressing the shortcomings of existing methods in utilizing image-based threat information. The code and the corresponding dataset will be released upon acceptance.

The mathematical capabilities of AI systems are complex and multifaceted. Most existing research has predominantly focused on the correctness of AI-generated solutions to mathematical problems. In this work, we argue that beyond producing correct answers, AI systems should also be capable of, or assist humans in, developing novel solutions to mathematical challenges. This study explores the creative potential of Large Language Models (LLMs) in mathematical reasoning, an aspect that has received limited attention in prior research. We introduce a novel framework and benchmark, CreativeMath, which encompasses problems ranging from middle school curricula to Olympic-level competitions, designed to assess LLMs' ability to propose innovative solutions after some known solutions have been provided. Our experiments demonstrate that, while LLMs perform well on standard mathematical tasks, their capacity for creative problem-solving varies considerably. Notably, the Gemini-1.5-Pro model outperformed other LLMs in generating novel solutions. This research opens a new frontier in evaluating AI creativity, shedding light on both the strengths and limitations of LLMs in fostering mathematical innovation, and setting the stage for future developments in AI-assisted mathematical discovery.