Stolen AI keys ran 500,000+ model calls in under eight hours and left one company with a bill of nearly $1 million

Cryptojacking used to nibble at your cloud bill for months. Its AI cousin can eat it in an afternoon.

What happened

Google's Mandiant incident responders say they handled several 2026 break-ins where criminals hijacked a company's access to AI models, a tactic called LLMJacking. Victims spanned finance, retail, tech and transport in North America, South America and Asia Pacific. The main way in was stolen or exposed cloud credentials: service account keys, GitHub tokens, session tokens. One gang scraped AI API keys from public GitHub repos and tested each with a request for a single token, too small to trip volume alerts.

The damage

Then the meter ran. In one case, an actor made more than 500,000 model calls in under eight hours with stolen keys to power Jet Night, a token resale service, and one victim was left with nearly $1 million in token costs. Jet Night sells an 'unlimited' plan for about $60 a month. Another group switched on AI services inside victims' cloud projects and passed $100,000 in charges. In several cases the victim had never even turned those AI services on; the attackers did. One actor also planted a backdoor built on the Manus agent framework, which asked a model what to do and ran the commands on the victim's machine.

How it was caught, and what happened next

Google expects LLMJacking to overtake cryptojacking. Its advice: spending caps, tightly scoped keys, secret scanning, and alerts the moment anyone enables an AI service.