Rising Threat of LLMjacking: Cybercriminals Exploit New AI Models for Illicit Gains
Emerging Threat: LLMjacking Targets DeepSeek Models Just Weeks After Launch
In a troubling development for cybersecurity, sophisticated "LLMjacking" operations have reportedly exploited stolen access to DeepSeek models, barely weeks following their public release. The term LLMjacking refers to the unauthorized exploitation of large language models (LLMs) such as those developed by OpenAI and Anthropic. By hijacking these powerful tools, criminals can run expensive tasks at the expense of unsuspecting account holders.
Recent research from Sysdig highlights alarming activity post the launch of DeepSeek’s DeepSeek-V3 model on December 26. Within days, LLMjackers had infiltrated the system, and in a similar fashion, access to the newly released DeepSeek-R1 was compromised the very next day after its January 20 launch. “This isn’t just a fad anymore,” states Crystal Morin, a cybersecurity strategist at Sysdig. “We are witnessing an escalation, far beyond our initial observations last May.”
At the core of LLMjacking lies the appeal of circumventing the potentially exorbitant costs associated with extended LLM usage. For instance, using GPT-4 around the clock can run up bills exceeding $500,000 per year. Attackers illicitly acquire cloud service credentials or application programming interface (API) keys, then employ reverse proxies to mask their activities while accessing these models.
The situation has led to flourishing underground communities on platforms like 4chan and Discord, enabling individuals to tap into these models for generating everything from school essays to explicit content. With the rise of LLMjacking, victims face considerable financial repercussions, as evidenced by one AWS user who encountered a staggering $730 bill after a mere few hours of compromised access. While AWS managed to assist this individual, the potential impact on enterprises could be catastrophic, marking a pressing concern in the landscape of cybersecurity.