1-minute read of a 4 min report

Google Tracks Stolen LLM Keys Behind Million-Dollar Cloud Bills

By

1-minute read

What a defender needs to know

Mandiant’s 2026 LLMJacking investigations found attackers using stolen cloud identities and model API keys to consume enterprise-paid inference. Fraudulent resource charges exceeded $100,000 in a single day in individual incidents, and cumulative charges reached $1 million at some organizations before access was revoked. One actor made more than 500,000 API calls in less than eight hours through a resale service, passing nearly $1 million in token costs to a victim organization.

Trace new model-service enablement and credential creation to the identities that made the changes. In AWS CloudTrail, correlate bedrock.amazonaws.com model-access operations with Converse and InvokeModel. In Azure Cognitive Services Activity Logs, check accounts/deployments/write and inspect inference traffic in Diagnostic Logs. For Google Cloud Agent Platform, pair ServiceUsage.EnableService with Data Access logs for PredictionService.GenerateContent; the relevant Data Access logs need explicit configuration. Tiny validation calls with max_tokens=1 can precede the bill spike.

Remove unneeded model access, rotate exposed keys, limit service accounts and scope API keys by model and environment. Set per-model quotas and budget alerts; confirm an authorized test workload produces the expected invocation and billing records and triggers the configured alert. If access was unauthorized, revoke keys and sessions and inspect newly created identities, SSH keys and secret-manager access. Google’s LiteLLM /key/regenerate rebind finding and the CVE Program’s CVE-2026-40217 record describe different endpoints; match gateway fixes to the behavior you are investigating.

Primary sources