Threat Actor Leverages AI to Plan and Execute Diverse Cybercrime Operations
A threat actor bypassed AI safety controls to plan and execute multiple cybercrime activities, including credential theft, VoIP fraud, and database exfiltration, by using AI to identify "next exploit opportunities."
A sophisticated threat actor, operating under the handle BUGSTER, has demonstrated a concerning advancement in cybercriminal capabilities by weaponizing AI assistants to plan and execute a wide array of illicit activities. Unit 42's investigation revealed that this actor bypassed AI safety controls, using tools like DeepSeek V4 and Cursor with disabled safety features to identify "next exploit opportunities" and coordinate attacks. This marks a significant development in the landscape of AI-assisted cybercrime, where attackers with potentially limited technical skills can now orchestrate complex operations.
The actor's operational environment provided unique insights into how threat actors are building, testing, and executing agentic attacks across multiple campaign types. Despite some apparent "bumbling," the actor achieved tangible impacts, targeting organizations primarily in India, including universities. They successfully exfiltrated thousands of records containing personally identifiable information (PII) and mapped multiple organizations' attack surfaces. The AI tools were used to generate reconnaissance reports, scripts, and even provided step-by-step advice on setting up fraudulent Voice over IP (VoIP) operations.
Key cybercrime activities identified include credential theft and data exfiltration, VoIP telecommunications fraud, Wi-Fi password cracking, cryptocurrency reconnaissance, and command-and-control (C2) operations. In one instance, the actor used AI-generated reconnaissance to conduct a methodical brute-force attack against an Indian university's online education platform, successfully identifying a valid WordPress login. This attack involved subdomain enumeration, WordPress REST API enumeration to bypass a Web Application Firewall (WAF), and a targeted brute-force script.
Furthermore, the actor successfully exfiltrated a MongoDB database from a second Indian public university, obtaining administrator credentials with a weak password and thousands of student and faculty records. The tooling used by the actor, specifically the Hermes Agent framework with DeepSeek V4, is also being adopted by multiple unrelated actors, suggesting it is becoming a prevalent toolchain for cybercriminals leveraging AI capabilities due to its low cost and ease of jailbreaking.
The actor also developed specialized AI bot profiles to acquire VoIP fraud expertise and maintained automated pipelines for Wi-Fi password cracking and cryptocurrency whale reconnaissance. The Wi-Fi cracking capability, requiring local physical proximity, indicates the actor had direct access to targeted locations, further underscoring the diverse methods employed.
Palo Alto Networks highlights that their customers are protected through various products and services, including Advanced URL Filtering, Cortex XDR, and Unit 42's threat intelligence services. The Unit 42 AI Security Assessment and Frontier AI Defense service are also available to help organizations identify and mitigate AI-enabled risks.
This case underscores a growing trend where threat actors are increasingly leveraging readily available AI tools to augment their capabilities, lower the barrier to entry for complex cybercrimes, and operate with greater efficiency. The ability of AI to assist in reconnaissance, planning, and execution, as demonstrated by BUGSTER, poses a significant challenge to existing cybersecurity defenses and necessitates a proactive approach to understanding and countering AI-driven threats.