AEGLIX News
Google Threat Intelligence Group and Mandiant report that LLMjacking attacks are generating fraudulent cloud AI charges exceeding $100,000 per day, with cumulative losses reaching $1 million at some organizations, according to research published Tuesday. Attackers compromise valid API keys and exposed endpoints to hijack large language model access, passing massive token costs to victim organizations.
Researchers identified two categories of threat actors driving the surge: operators scanning for publicly exposed LLM endpoints and actors stealing valid API keys from unprotected machines. Compromised credentials include cloud service account keys, GitHub tokens, active session tokens, and LLM API keys. Threat cluster UNC6993 uses gateway tools to aggregate stolen keys and distribute requests, likely selling access through underground forums.
Attackers monetize hijacked access through services called LLM mixers, which relay subscribers' prompts through stolen API keys. One mixer, Jet Night, sells unlimited access plans for $60 per month. In a documented case, a threat actor executed more than 500,000 LLM API calls in under eight hours using stolen third-party keys to power jet-night.com, passing nearly $1 million in token costs to the victim. Request volumes through hijacked access can spike suddenly, sometimes exceeding 1,000 requests per second. Researchers also observed exploitation of CVE-2026-40217, a LiteLLM vulnerability allowing arbitrary code execution via the custom-code guardrail testing endpoint.
Google and Mandiant continue tracking UNC6993 and associated LLMjacking infrastructure. The research highlights the financial severity of credential theft targeting AI workloads and the role of underground markets in monetizing compromised access. Organizations are advised to secure API keys, monitor for anomalous request volumes, and protect cloud credentials and service accounts from exposure.
Report prepared 2026-11-05 · Based on reporting from BleepingComputer. Source publication date above; subsequent developments may not be reflected.
Sources
Read original source · BleepingComputer (opens in a new tab)