This research reveals how malicious actors can exploit Code LLMs to generate conventional malware through backdoor manipulation and adversarial instruction attacks.

• Identifies a novel attack vector where backdoored Code LLMs can be manipulated to produce traditional malware
• Demonstrates that attackers can use adversarial instruction tuning to trigger hidden vulnerabilities
• Shows that these attacks can bypass standard security screening methods
• Exposes a critical intersection between emerging AI capabilities and traditional cybersecurity threats

This research highlights urgent security implications as Code LLMs become increasingly integrated into software development workflows, requiring new defensive approaches that span both AI and traditional security domains.

Double Backdoored: Converting Code Large Language Model Backdoors to Traditional Malware via Adversarial Instruction Tuning Attacks