A new threat vector known as LLMjacking is rapidly emerging as cybercriminals pivot from traditional cryptojacking to hijacking enterprise artificial intelligence infrastructure. According to a recent report from Google's Threat Intelligence Group, malicious actors are systematically stealing API keys, cloud credentials, and service account tokens to siphon expensive large language model resources. Unlike traditional resource theft that accumulates charges slowly over time, compromised AI endpoints can rack up over $100,000 in billing charges within mere minutes as attackers leverage high-volume inference streams and premium foundation models.
The mechanics of these attacks typically begin with hackers gaining an initial foothold through publicly exposed LLM endpoints or compromised cloud identity tokens, such as GitHub Personal Access Tokens and active session sessions. Once inside a victim's cloud project, threat actors work to escalate privileges and establish persistence before executing the endgame. This involves unauthorized hardware provisioning, token proxying, and automated model streaming. To bypass standard security rate limits, attackers frequently employ a burn-and-rotate strategy by continuously creating and deleting short-lived service accounts while closely monitoring the cloud provider's financial interface.
Investigators have uncovered a maturing underground economy built entirely around this specialized crime. The ecosystem features a distinct division of labor, where certain hacking groups specialize in securing initial enterprise access while specialized operators, dubbed LLM Mixers, manage the downstream monetization infrastructure. These mixers utilize open-source API gateways and custom relay infrastructure to pool stolen credentials, meter usage, and resell unauthorized AI access to underground subscribers. Google urges organizations to treat unauthorized enablement of LLM services as a critical security event, strictly enforce the principle of least privilege, and increase overall visibility across AI deployment platforms to prevent catastrophic financial loss.
Comments