Don’t Let a Refusal Count as a Pass: Defending AI Malware Scanners Against Refusal-Based Evasion
Supervisor Name
Yousef Sabbah
Supervisor Email
ysabbah@birzeit.edu
University
Birzeit University
Research field
Computer Science
Bio
Dr. Yousef Sabbah is an Associate Professor of Cybersecurity at Birzeit University, Palestine. He holds a PhD from Cairo University and is a national leader in digital transformation. In 2020, he authored Palestine’s first National Digital Transformation Policy including cybersecurity, approved by the ITU and the Palestinian Cabinet. He received the Palestine Scientist Prize (2017) and the Scientific Research Excellence Prize (2020). His expertise spans network security, cryptography, ISO quality management, and AI applications. Dr. Sabbah has consulted for the UN (ESCWA) and led multiple EU-funded projects. He supervises master’s research on AI, blockchain, and digital governance, and has published extensively on cybersecurity, e-learning, and smart systems.
Description
AI-based malware scanners use a large language model (LLM) to read code or software packages and judge whether they are malicious. A new evasion technique targets this step; attackers hide fake text about banned topics inside malicious files, so the scanner's safety rules make it refuse to continue, never reaching the real malicious payload. JFrog reported this in a Shai-Hulud malware sample in June 2026, and ESET described a similar case it calls GuardBreaker. Traditional static analysis is not fooled by this trick, so combining it with an AI scanner could work better than either alone. Little is known about how often local models fail this way or how well proposed fixes work. This project measures the problem on four different local models under five different bait conditions, using explicitly harmless test files and no real malware, and then tests multiple defenses on the models to make them treat a refusal as a warning instead of a full pass. The result is a public benchmark dataset, a workshop-ready report, a one page mitigation guide (in English and Arabic) and an open-source pipeline that security teams can use to make their AI-based scanners more reliable.