Stolen AI API keys generated nearly $1M in charges for a single company

LLMJacking is the theft of cloud and AI-model access to run unauthorized workloads and sell model access, often producing far faster and larger bills than cryptojacking. Investigations in 2026 found incidents across several regions and industries, with fraudulent usage exceeding $100,000 in a day and, in one case, nearly $1 million charged to a company. Attackers commonly entered through exposed or stolen cloud credentials - such as service-account keys, IAM credentials, GitHub tokens, and active sessions - or API keys leaked in public code repositories. Some tested stolen keys with tiny requests to avoid triggering usage alerts; others exploited publicly exposed model endpoints.

After gaining access, attackers searched repositories, cloud services, logs, databases, and secret stores for more credentials, then expanded privileges and established persistence through unauthorized accounts, policies, SSH keys, or compromised virtual machines. Some also exploited a LiteLLM gateway flaw to obtain administrative control. They monetized access by generating new keys, enabling premium models, running high-volume inference, provisioning GPU instances, or routing stolen quotas through proxy services. A growing ecosystem separates initial-access operators from “LLM Mixers,” which pool credentials and resell access while concealing its origin. One service made more than 500,000 API calls in under eight hours and passed nearly $1 million in token costs to its victim. The findings show that securing AI access requires protecting cloud identities and keys, limiting permissions, monitoring usage and billing spikes, and restricting exposed endpoints.