Attackers Hijack LLM Services to Generate $1M in Fraudulent Cloud Charges

Attackers compromise enterprise LLM infrastructure via exposed endpoints and stolen credentials, then resell access through LLM mixers—services offering subscribers unauthorized LLM use at steep discounts via proxy infrastructure. Request volumes spike sharply, sometimes exceeding 1,000 per second, while the victim organization gets billed for the full load. Google researchers documented cases where organizations incurred over $100,000 per day in fraudulent charges, with cumulative costs reaching $1 million. One attacker generated nearly $1 million in charges using 500,000+ API calls over eight hours.

The attack ecosystem relies on two threat actor types: those scanning for exposed endpoints and those stealing credentials from unprotected systems. Services like Jet Night charge as little as $60/month for unlimited access. Attackers gain initial entry through compromised cloud credentials, service account keys, GitHub tokens, and exploits like CVE-2026-40217 in LiteLLM. Once inside, they search for additional secrets in logs and databases, establish persistence, and sometimes use the victim’s infrastructure to host proxy endpoints. Google recommends strict IAM controls, centralized secrets management, least-privilege access, phishing-resistant MFA, shadow AI monitoring, spending caps, and comprehensive audit logging.