Research
Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
2,033 items
Researchers from multiple countries ran a cross-cultural experiment to examine how cultural orientations relate to phishing victimization and phishing recognition. The source text provided contains only publication metadata (November 2026, Computers & Security, Volume 170, and the author list), without the study's method or findings.
The paper, published in Computers & Security (Volume 170) in November 2026 by Chong Lei and Panpan Fu, presents CleanStack, a new dual-stack design. The source text provided gives only the title, publication details and authors, so the research question, method and findings cannot be stated from it.
The source reports a 2026 Computers & Security paper by Shuai Liu, Yun Zhong, Xiangxu Meng, Wenfeng Li and Kanglian Zhao. Only the bibliographic details are given: publication in November 2026, Volume 170, and the title M-LITO, a location imitation method against offloading- and RSSI-based side-channel inference in maritime edge networks. The source text does not state the research question, method or findings.
The source is a November 2026 Computers & Security paper (Volume 170) by Kyriakopoulos, Kiarie, Aristodemou and colleagues. The provided text contains only publication metadata: date, journal, volume and author list. It states no research question, method or findings, so no summary of the work can be given from it.
Tong Wu, Qingyu Ou, Fang Luo, Weidong Lv, Wei Fu and Dawei Zhou published HMM-MET in Computers & Security, Volume 170, in November 2026. The source text provided contains only the publication metadata and author list, so the research question, method and findings cannot be summarized from it.
The source reports on real-time microburst attack detection using sketch-based queue monitoring in programmable switches, published in Computers & Security, Volume 171, by Lilima Jain and Venkanna U in December 2026. The source text provided contains only bibliographic metadata and no abstract, method details, or findings, so the research question, method and results cannot be summarized from it.
Cai and colleagues study formal security properties of CAS, a system whose authors are not named in the source text beyond the listed authors. The source text provided contains only publication metadata (Computers & Security, Volume 171, December 2026) and author names, with no description of the research method or findings.
The source reports on a study of phishing detectors and how well they hold up under shifts caused by generative AI. It was published in Computers & Security, Volume 171, in December 2026, by Oumkaltoum Arharbi and Abderrahim El Qadi. The provided source text contains no further details of the method or findings.
DifAttack++ is a score-based black-box adversarial attack that works in a hierarchical disentangled feature space. It splits an image's latent representation into an Adversarial Feature and a Visual Feature using a cross-domain autoencoder, then optimizes only the Adversarial Feature using query feedback from the victim model. The authors report higher Attack Success Rate and query efficiency than state-of-the-art methods, with comparable visual quality.
Researchers propose SIC-AE, a framework that reduces hallucination in large language models by running self-correcting loops on each component model, then ranking candidate responses and fusing the top ones through an autonomous weight update mechanism. Built on low-parameter open-source LLMs, it reports accuracy gains of 1-5% over state-of-the-art self-refinement techniques and a further 1-8% from adding the ensemble across three benchmarks covering factual and logical reasoning questions.
Researchers introduce TUBench, a benchmark that tests Large Vision-Language Models on unanswerable questions built from natural images, code snippets, geometry diagrams and statistical tables. They evaluated 39 foundational models, and GPT-5 reached the highest average accuracy at 83.2% in judging answerability. Manual analysis found GPT-5 gave both a correct answer and explanation in only 63% of cases, with hallucinations causing 87.9% of its incorrect explanations.