Google, Mandiant detail LLMJacking AI model hijacking

Google Threat Intelligence Group and Mandiant published a blog analysis on November 4, 2026 of an attack technique called LLMJacking, in which attackers hijack and resell access to companies' AI models, according to gamebusiness.jp.

In the technique described in the analysis, attackers steal API keys and other credentials by gathering them from public GitHub repositories or by hijacking sessions, then use them to enter the target's cloud environment. The stolen keys are pooled behind illicit proxy gateways and sold on underground marketplaces as discounted access to frontier models.

In one investigation, fraudulent charges exceeded $100,000 per day before the unauthorized access was stopped, and cumulative damage reached $1 million. In another case, a token routing service offered buyers unlimited frontier-model access for about $60 a month while the victim company was billed about $1 million in model usage costs.

Another attacker made more than 500,000 LLM API calls in under eight hours, using a method described as "Burn and Rotate" — creating and deleting short-lived service accounts to evade rate limits — and checked cloud billing screens to gauge the cost impact.

Victims include organizations in financial services, retail, technology and transportation in North America, South America and the Asia-Pacific region. Recommended countermeasures include detecting shadow LLMs, setting fine-grained API key scopes, and automated scanning for sensitive information.

Earlier attacks on compute

Cryptocurrency mining attacks that steal computing resources have existed for some time, but LLMs are costly per call, so bills can balloon quickly.