Malware developers add nuclear and bioweapon phrases to spyware

Malware developers are exploiting the LLM Safety Refusal mechanism to obstruct security analysis. They insert text related to nuclear or biological weapons into spyware code, so that when an AI-based security scanner attempts to analyze the file, the model refuses to perform the analysis itself, citing safety policy.

This type of attack reveals a security vulnerability that arises when there is excessive reliance on a model's primary safety alignment. Attackers identify the model's refusal conditions and exploit them as a secondary blind spot, a threat that can apply to both closed models and open models.

The key issues and directions for response are as follows:

• Importance of intent judgment: Security analysis pipelines should be designed not through simple keyword matching, but by grasping the actual intent of the text in order to avoid prompt injection.
• Model dulling problem: In systems dealing with complex cybersecurity problems, care is needed so that excessive safety features do not degrade the model's analytical performance.
• Various insertion methods: Beyond plain text, methods of hiding such phrases in white-colored text, image watermarking, and PDF metadata are also being discussed.