Google reports AI access theft driving $1 million cloud charges

Google Threat Intelligence Group and Mandiant say attackers are hijacking enterprise access to large language models and reselling it, leaving some organizations with fraudulent cloud charges exceeding $100,000 in a day. In some investigated incidents, cumulative charges reached $1 million before access was revoked, the researchers reported. The activity, known as LLMjacking, targets access to AI models and services rather than merely using victims’ computing resources. Attackers gain entry through exposed or stolen cloud credentials, API keys, or publicly accessible model endpoints, then make requests that cloud providers bill to the affected organization. Researchers described a growing division of labor: some actors obtain access, while others operate LLM Mixers, services that route and resell unauthorized model access to paying customers. The mixers can use stolen credentials through proxy infrastructure, making the requests appear as legitimate consumption by the victim’s account. In one case, an actor used stolen third-party keys through the Jet Night routing service to make more than 500,000 LLM API calls in under eight hours. Google researchers said the activity...