Token thieves costing companies $100K+ per day in illicit LLM charges
Reported by The Register1 source ↓
Google’s Mandiant incident responders report that criminals are hijacking companies’ LLM access to run up charges and monetize stolen API keys. The activity, known as LLMJacking, has become more organized.
Key points
- Mandiant says some victims accumulated $1 million in charges before cutting off attackers
- Google documented criminals stealing API keys, creating new keys and enabling cloud services
- A token thief made more than 500,000 API calls in under eight hours, Google reports
- Google identified separate roles for gaining access and monetizing compromised LLM services
- Mandiant recommends tighter access controls, centralized secrets management and automated key scanning
Details
The Register reports that Google’s Mandiant responders found criminals using hijacked corporate LLM access to consume AI services and generate illicit charges. In one case, the group Google tracks as UNC6993 obtained LLM API keys, created additional keys and enabled cloud services, resulting in $100,000 in charges. Mandiant says some organizations’ cumulative charges reached $1 million before access was revoked.
Google also documented an intruder making more than 500,000 API calls in under eight hours with stolen third-party keys to operate a token-routing service called Jet Night. Google assessed that the service likely relied on stolen keys because it offered access to advanced models at prices below typical market rates. The Register notes that Google pointed to Jet Night outage reports as likely resulting from victims discovering compromised keys and revoking them.
Google describes a division of labor between criminals who obtain access to cloud and AI services and operators of downstream monetization infrastructure, which it calls “LLM Mixers.” In a mid-2026 incident, UNC6993 used a misconfigured service-account key to access a cloud instance and then gain model API access; a session-hijacking page also tricked users into submitting active access tokens. Mandiant recommends limiting permissions, rotating secrets and scanning code for exposed keys.
Written from coverage as of
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh