Google’s November 2026 research describes attacks that hijack access to enterprise AI models and generate extraordinary compute charges.
The new target
Google Threat Intelligence Group and Mandiant described 2026 incidents in which intruders used compromised access to enterprise large language models. The researchers called the abuse LLMJacking: malicious consumption of costly model resources for the attackers’ benefit.
The research reported individual incidents with fraudulent usage bills exceeding US$100,000 in a single day, and some organizations facing cumulative charges approaching US$1 million before access was cut off. These are researcher-reported case outcomes, not a claim that every exposed key leads to losses on that scale.
Why this is different from ordinary credential theft
A stolen API key or cloud credential is valuable even without downloading a database. The attacker may burn through paid compute capacity, disrupt quotas, or conceal activity among legitimate inference calls. Cloud billing then becomes an important signal alongside access logs.
A defender’s checklist
Inventory model endpoints and service accounts; restrict keys to the smallest needed scope; impose quotas and spending alerts; rotate credentials after suspected exposure; review anomalous usage by time, location and workload. Logs should distinguish approved deployment traffic from strange consumption spikes.
The bigger implication
AI products are part of ordinary infrastructure now. They need the same identity governance, monitoring and incident response as storage, databases and remote-access tools. Security cannot stop at model safety prompts when the surrounding cloud account is exposed.
Sources and further reading
IPCRIME research and analysis based on the linked primary or attributed sources; no independent incident-response access is claimed.