A paper by Prof. Bogdan Księżopolski and Leszek Wroński from the Department of Cybersecurity at Kozminski University has been accepted for SECRYPT 2026 – the 23rd International Conference on Security and Cryptography. The research addresses one of the most pressing challenges in cybersecurity today: protecting systems that use large language models (LLMs) against emerging forms of attacks.
The paper, titled “LogSanitizer: Defending LLM-Integrated SOCs against Backdoor Triggers Delivered through Firewall Logs”, was accepted as a Short Paper and will be presented during SECRYPT 2026 in Porto. The authors examine the security of Security Operations Centers (SOCs), which increasingly rely on language models to analyze events, logs, and potential incidents. The study focuses on a scenario in which the LLMs themselves become a source of risk because they have previously been poisoned and contain hidden mechanisms capable of triggering backdoor attacks.
The key contribution of the paper is the proposed protection module that sanitizes incoming data before it reaches the poisoned model, preventing the hidden backdoor mechanism embedded in the LLM from being activated.
This is an important research direction because the rapid development of tools based on generative artificial intelligence is changing not only the way cybersecurity teams operate, but also the range of threats that must be considered when designing secure systems. Research conducted at Kozminski University contributes to the global debate on LLM resilience, input data security, and the responsible deployment of AI in organizational environments.
SECRYPT is an international conference dedicated to information and communication security, privacy, data protection, and applied cryptography. Among the thematic areas of the 2026 edition are AI and machine learning security.
The cybersecurity of LLMs is no longer solely a research topic. Today, it concerns critical systems that support event analysis and decision-making in organizations. That is why we must study not only the capabilities of the models themselves, but also the ways in which they can be misled by the data they process,” says Prof. Bogdan Księżopolski from the Department of Cybersecurity at Kozminski University.
