An adversarial evaluation framework for LLM-integrated Security Operations Centers.

RedSOC is an open-source framework that systematically evaluates how AI-powered security assistants fail under adversarial conditions — and whether detection methods can catch those failures.

As two-thirds of organizations now deploy AI in their SOC environments, no unified framework exists to red-team these systems. RedSOC fills that gap.

• Prompt Injection (direct, indirect, multi-turn)
• RAG Poisoning (corpus poisoning, backdoor attacks)
• Multi-Agent Hijacking
• Concept Drift under adversarial conditions

• SOC-RAG Pipeline — simulated AI security assistant
• Attack Simulator — benchmarked attack modules
• Detection Layer — semantic anomaly + provenance tracking
• Benchmark Runner — automated results and charts

🚧 Active development — April 2026

Paper: https://doi.org/10.6084/m9.figshare.32016498 Benchmark Dataset: https://doi.org/10.6084/m9.figshare.32016534 Benchmark Charts: https://doi.org/10.6084/m9.figshare.32016564 Zenodo Record: https://zenodo.org/records/19519927

If you use RedSOC in your research, please cite:

Krishnakaanth Reddy, Y. (2026). RedSOC: An Adversarial Evaluation Framework for LLM-Integrated Security Operations Centers. arXiv preprint.

Krishnakaanth Reddy — IEEE Professional Member | ACM Professional Member | AI Security Researcher

MIT License