Google Threat Intelligence Group and Mandiant have uncovered a growing cybercriminal operation known as LLMjacking, where attackers steal enterprise AI credentials and resell access to expensive large language models (LLMs). During investigations in 2026, organizations across financial services, retail, technology, and transportation experienced fraudulent cloud charges exceeding $100,000 in a single day, with some incidents accumulating approximately $1 million. Attackers exploited stolen API keys, compromised cloud accounts, and exposed AI services to generate massive volumes of unauthorized requests.
CYBERCRIMINALS RESELL STOLEN AI ACCESS THROUGH LLM MIXERS
Unlike traditional cryptojacking, which uses compromised computing resources to mine cryptocurrency, LLMjacking exploits commercial AI services and cloud infrastructure. Criminals operate underground proxy platforms called LLM mixers, allowing customers to access premium AI models using stolen enterprise credentials. Google observed one operation generating more than 500,000 API requests in under eight hours, resulting in nearly $1 million in token costs. Attackers also targeted exposed Ollama endpoints, GitHub tokens, and cloud service accounts while exploiting vulnerabilities such as CVE-2026-40217 in LiteLLM gateways to gain administrative privileges.
ORGANIZATIONS URGED TO STRENGTHEN AI CLOUD SECURITY
Security researchers warn that LLMjacking can escalate beyond financial fraud into broader cloud infrastructure compromises. Attackers have created unauthorized service accounts, activated previously unused AI services, deployed GPU instances, and established persistent access to compromised environments. Organizations should enforce strict API request quotas, spending limits, least-privilege permissions, and phishing-resistant multifactor authentication. Security teams should also monitor unusual AI model requests, unauthorized service activation, and sudden consumption spikes while promptly revoking compromised credentials and patching vulnerable AI gateways.