越獄版 DeepSeek V4 協助黑客透過人工智能網絡攻擊竊取逾 5,200 份印度大學記錄

Unit 42's investigation into the actor known as BUGSTER details how AI models are enabling less-skilled individuals to execute complex cybercrime. The attacker configured the Hermes Agent framework with a jailbroken DeepSeek V4 as its reasoning engine and used Cursor IDE, an AI-assisted development tool. This combination allowed the generation of scripts and reconnaissance reports, compensating for the attacker's limited traditional technical expertise.

The actor successfully targeted multiple Indian organizations. AI-generated reconnaissance was used against an Indian university's online education platform to identify subdomains and extract user information, which led to credential attacks. In a separate incident, an administrator account at another public university was compromised, resulting in the exfiltration of a MongoDB database containing over 5,200 personally identifiable student and faculty records.

The case also illustrates AI's role in diversifying criminal activities. The hacker created specialized AI profiles, such as "alice" for VoIP fraud advice and "pandarouter" for carrier research. This allowed a single actor to operate across diverse attack types, including credential theft, data exfiltration, VoIP fraud, Wi-Fi password cracking, and cryptocurrency reconnaissance, activities that would traditionally demand distinct specialist skill sets.

Unit 42 noted the actor repeatedly tested ways to bypass AI safety controls, eventually settling on DeepSeek V4 due to its minimal safety friction. Researchers also found other actors adopting this Hermes and DeepSeek combination, suggesting its growing appeal to cybercriminals because of its low cost and limited safety friction.