Researchers from Abu Dhabi’s Khalifa University, in collaboration with the University of Bonn and the University of Milan, have developed RedSage, a groundbreaking open-source large language model (LLM) designed specifically for cybersecurity applications. The project’s research paper has been accepted for presentation at ICLR 2026, a premier conference for representation learning, highlighting the UAE’s growing influence in frontier AI research.
Addressing a Critical Security Gap
RedSage was created to solve a major dilemma for organisations: the choice between using proprietary AI services that risk exposing sensitive data to external APIs, or relying on open-source models that often lack specialised cybersecurity training. RedSage addresses this by enabling secure, on-premises deployment on consumer-grade GPUs, ensuring privacy-preserving operations.
The 8-billion parameter model is designed to support a wide array of security workflows, from frameworks like MITRE to offensive techniques and tool usage, without compromising data integrity.
Performance and Innovation
Despite its relatively small size, RedSage has demonstrated superior performance, surpassing baseline models by up to 5.59 points on key cybersecurity benchmarks. The instruction-tuned variant even outperformed the much larger Qwen3-32B model, proving that domain-aware pretraining can yield better results than general-purpose models with more parameters.
To achieve this, the research team curated 11.8 billion tokens of cybersecurity-focused data and developed an innovative agentic augmentation pipeline. This pipeline simulates expert workflows to generate 266,000 high-quality, multi-turn samples for supervised fine-tuning. The team also introduced RedSage-Bench, a rigorous new benchmark with over 30,000 questions to evaluate the model’s technical proficiency.
Commitment to Open-Source Research
In a move to accelerate open research and reproducibility, the RedSage team is releasing all models, datasets, and code to the public. The model family includes three variants to suit different needs: RedSage-8B-Base for further fine-tuning, RedSage-8B-Ins for multi-turn chat and explanations, and RedSage-8B-DPO, a production-ready assistant. This open approach directly addresses the limitations of previous projects that often withheld crucial data and development pipelines.
A Hub for Specialised AI
The launch of RedSage is part of Khalifa University’s broader strategy to develop specialised AI models for knowledge-intensive technical domains. The university’s 6G Research Centre has previously released telecom-specific LLMs, such as TelecomGPT, in collaboration with the global industry association GSMA. These initiatives underscore the university’s commitment to advancing specialised AI to support strategic national goals, including the UAE’s target for 6G implementation by 2030.
About Khalifa University
Khalifa University of Science and Technology is a world-class, multi-disciplinary, research-intensive university in Abu Dhabi, UAE. It has three colleges, three research institutes, 18 research centers, and 17 departments covering a broad range of disciplines in science, engineering, and medicine. The university is dedicated to providing an unparalleled experience in education and research that prepares leaders and critical thinkers to address the world’s most pressing challenges.
Source: Middle East AI News
