Skip to content
nulltapE-reader edition

← All e-reader articles

4 min read

Google Tracks Stolen LLM Keys Behind Million-Dollar Cloud Bills

Mandiant traced stolen cloud and model credentials into unauthorized inference, reseller proxies and rapid charges. Hunt new model access and meter usage before bills surge.

Read on the standard site

A translucent access key lodged in a cloud-compute meter with a bright consumption gauge.

An enterprise can lose control of its model bill before its usual cloud-spend alarms tell a convincing story. Mandiant investigated multiple LLMJacking intrusions in 2026 across financial services, retail, technology and transportation organizations in the Americas and Asia Pacific. Attackers used stolen cloud identities, service-account keys and LLM API keys to run models on someone else’s account. Mandiant saw fraudulent resource charges exceed $100,000 in a single day in individual incidents; cumulative charges at some organizations reached $1 million before access was revoked.

The mechanism reaches beyond a leaked developer key. One actor made more than 500,000 LLM API calls in less than eight hours through a token-routing service called jet-night[.]com, passing nearly $1 million in token costs to the victim organization, according to Google Threat Intelligence Group (GTIG). The team says its evidence points to stolen keys powering that service. For a cloud operator, the first useful question is which identity enabled model access, created fresh credentials or started paying for inference.

Stolen identities open model accounts

Mandiant saw attackers arrive with IAM credentials, service-account keys, GitHub personal access tokens and active session tokens. In one incident, UNC6993 used a misconfigured service-account key to gain access to a cloud instance, then granted itself access to an LLM API. In another, a compromised federated identity created fresh access keys. A separate actor scraped exposed corporate LLM API keys from public repositories and tested them with tiny requests marked max_tokens=1.

A max_tokens=1 key test can resemble routine, low-volume model traffic before the large invoice arrives. GTIG describes the underlying business as “LLM Mixers”: gateways that pool compromised access, proxy requests from paying subscribers and meter the stolen quota. The compromised account pays for the upstream model calls; the reseller collects from the downstream customer. The recorded service newapi[.]pro is another reported resale destination. Use both attacker-linked domains as investigation indicators in proxy and gateway logs.

Privilege changes extend the theft

After entry, the actors searched repositories, cloud services, logs and secret stores for more credentials. Mandiant observed new IAM users and service accounts, administrative policy attachments and SSH keys injected into compute instances or instance metadata. Some enabled model services that the organization had previously left inactive. Others created short-lived service accounts to distribute requests and watched billing while high-volume inference ran.

One follow-on cluster used prevalidated LiteLLM virtual keys and abused a user-ID rebind in the gateway’s /key/regenerate endpoint to obtain PROXY_ADMIN credentials, GTIG reports. Treat that as an intrusion finding when reviewing gateway permissions. The CVE Program record for CVE-2026-40217 assigns that identifier to code execution through /guardrails/test_custom_code, a different LiteLLM behavior. Match the observed endpoint and vendor advisory to the deployed build before choosing a gateway fix.

Model usage has three log paths

Look for a change in model access before the usage spike. In AWS CloudTrail, GTIG recommends watching bedrock.amazonaws.com for CreateFoundationModelAgreement and PutUseCaseForModelAccess, then correlating Converse, InvokeModel and InvokeModelWithResponseStream with the principal that enabled access. New access keys and users near those calls deserve the same investigation. For Azure, inspect Activity Logs for Microsoft.CognitiveServices/accounts/deployments/write and Diagnostic Logs for inference requests: the management and data planes produce different evidence.

In Google Cloud, alert on google.api.serviceusage.v1.ServiceUsage.EnableService for Agent Platform services and review Data Access logs for google.cloud.aiplatform.v1.PredictionService.GenerateContent and google.cloud.aiplatform.v1.PredictionService.StreamGenerateContent. GTIG notes that Gemini Enterprise Agent Platform Data Access logging is disabled by default because of volume and ingestion cost. Enable the relevant IAM audit policy for the projects you need to investigate; otherwise those model calls will be absent from Data Access logs.

Limit the account and test it

Inventory the identities and API keys that can enable or invoke paid models. Remove unused model access, reduce service-account permissions and scope model keys by endpoint, environment and usage limit. Rotate exposed keys from code repositories and investigate the issuing identity and any credentials created after exposure. Restrict publicly reachable model endpoints, including Ollama deployments, because GTIG also observed attackers testing exposed endpoints to run their own models.

Set per-model API quotas and granular budget alerts, then confirm that the control actually triggers on an authorized test workload. Compare the resulting billing and invocation data with the cloud audit operations above. Start with unauthorized service enablement, low-volume max_tokens=1 validations and new credentials, then follow their model calls and cost. When activity is unauthorized, revoke the affected keys and sessions, remove attacker-created identities and SSH keys, and inspect secret-manager access for secondary exposure. A bill can rise within hours; the identity change that opened the meter is the earlier signal.

Primary sources