AIDEFEND in Action AIA-0147 Incident

GTIG / Mandiant: Multi-Cloud LLMJacking and "LLM Mixer" Monetization

Google Threat Intelligence Group and Mandiant reported LLMJacking across several 2026 investigations: actors used other organizations' credentials to run cloud language models at the victims' expense. Some incidents incurred more than US$100,000 in a day. A separate token-routing service handled over 500,000 calls in less than eight hours, passing nearly US$1 million in costs to one victim. The investigations show several routes into cloud or AI gateways and different ways to cash out. For LiteLLM operators, v1.83.14-stable includes the fix for the cited CVE but still lacks the key ownership-change check added in v1.84.0.

Resource HijackingCredential TheftCredential GovernanceCloud IAMVertex AI
9 applicable AIDEFEND defenses
Source: The Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model Resources 
Authors: Christopher Romano, Codi Starks, Garrett Mahoney, John Wolfram, Josh Murchie, K Singh, Mathew Potaczek, Mike Lombardi, Rik Dolfing (Google Threat Intelligence Group)
Original article: Nov 2, 2026

Threat Analysis

  • Entry points differ between investigations. Mandiant describes cases in North America, South America and Asia-Pacific. Actors used stolen IAM credentials, cloud service-account keys, GitHub personal access tokens (PATs) and model API keys. One actor checked leaked model keys with very small output limits, an indicator such as repeated max_tokens=1 calls rather than proof by itself. Other cases involved an exposed Ollama model service or UNC6993 phishing for an active cloud session. One case combined a stolen GitHub PAT with a suspected web application vulnerability in a containerized cloud environment.
  • What investigators observed after entry. Across different engagements, actors discovered models and regional quota, searched repositories and data-warehouse metadata for credentials, and abused cloud service-account token issuance to access secret stores. Some left misleadingly named billing-optimization accounts or added SSH keys. One cloud intrusion used publicly reachable proxies under the D1337-KAIROS persona and deployed a containerized backdoor using the Manus agent framework. The backdoor queried an upstream LLM to reason through instructions and executed the resulting commands locally, supported by SearXNG and a sandbox execution environment. These observations do not supply intermediate steps for the Jet Night case.
  • The LiteLLM escalation and its fixes need separate checks. GTIG reports that a follow-on cluster acquired pre-validated Virtual Keys from the initial-access actor and exploited missing authorization for an ownership change through /key/regenerate to obtain PROXY_ADMIN privileges. GTIG cites CVE-2026-40217, a public vulnerability identifier. The maintainer advisory, public CVE record and CISA KEV entry added on November 18, 2026 instead identify a sandbox escape at /guardrails/test_custom_code. Public records establish this mismatch, but do not independently establish the exact vulnerability used in that incident.

    LiteLLM v1.84.0 adds the ownership-change authorization check and repairs the Enterprise license-check bypass in key regeneration. An unmodified v1.83.14-stable deployment already includes the cited CVE fix, but still lacks this ownership-change check. Operators therefore need to verify both repairs separately.

  • How the stolen capacity becomes revenue. A pay-as-you-go case involved premium inference and service-account churn to bypass rate limits. GTIG describes AWS Bedrock, Google Cloud Vertex AI and Azure OpenAI in the wider activity. The report calls aggregators LLM Mixers because they combine access from multiple sources behind a single resale interface. It associates UNC6993 with new-api and gcli2api, and describes LiteLLM gateways in related investigations. Jet Night used stolen third-party keys to route over 500,000 calls in less than eight hours, at nearly US$1 million cost to one victim. A separate example routed access to newapi.pro.

Applicable AIDEFEND Defenses (9)

AID-H-004.002
Service & API Authentication
Very High
Replace independently usable static model keys with provider-native workload authentication where supported. Bind the authorized inference identity to approved ways of reaching the provider and verify that unapproved direct invocation is denied. In AWS, model-runtime access and VPC endpoint controls must be evaluated together; a private endpoint by itself does not remove public API authorization. Apply independent service authorization before LiteLLM. For example, use Istio mutual TLS and route policy to limit inference clients to approved inference routes, with key administration requiring a separate trusted identity. Verify that all listeners apply the policy and that inference clients cannot reach the backend directly. For the exposed Ollama service example, keep the unauthenticated local API on a private listener and expose only an authenticating ingress. Require scoped workload tokens and deny every uncovered route before invoking Ollama. Test that an authorized inference client can still use the service and that anonymous requests and direct backend access fail.
AID-E-001.003
AI Agent & Workload Principal and Issuance Disablement
Very High
Disable the precise attacker-created workload principals at the authoritative cloud identity source and block fresh credentials. Remove unauthorized bindings after preserving the account and usage evidence. Verify that an approved test of fresh issuance for each affected principal is denied. Handle existing tokens, stolen keys and SSH access as separate response actions.
AID-H-003.010
Deployed AI Software Vulnerability Remediation Lifecycle
Very High
Deploy a trusted, supported LiteLLM release containing the ownership-change authorization check introduced in v1.84.0. The v1.83.14-stable CVE fixes do not include this check, which rejects unauthorized changes to the key owner during regeneration. Verify the running version or image digest on every replica that can handle key management requests, and confirm that allowed key maintenance still works with a test key while the ownership policy is enforced.
AID-I-003.003
Resource, Token & Spend Budget Throttling
High
Meter inference centrally and deny new calls before they exceed a hard quota for the verified project, workload and model. Count consumption across changing service accounts so rotation cannot reset the budget. An external gateway ledger protects only the routes forced through it. Configure a genuine cap only where the product supports one and validate it at the effective enforcement point.
AID-D-011.004
Non-Human Identity & Delegated Token Abuse Detection
High
Correlate cloud provider model-invocation logs (AWS Bedrock CloudTrail data events, GCP Vertex AI audit logs, Azure OpenAI diagnostic logs) with IAM service-account creation, token impersonation (getAccessToken), secret-manager bulk reads, and multi-region model activation to detect direct-to-provider LLMJacking and "burn-and-rotate" service-account patterns.
AID-E-002
AI Runtime Process, Lease & Pod Eviction
High
Quarantine and stop the confirmed unauthorized Manus agent runtime using the responsible cloud execution service or host supervisor. Verify the run is absent and contain retries. Separately evict the backing image, reverse proxy, principals or secrets when the investigation confirms their involvement.
AID-H-003.005
Infrastructure as Code (IaC) Security Scanning for AI Systems
Medium
Where the repository contains infrastructure or deployment configuration for AI workloads, block hardcoded API keys in source, generated manifests and rendered plans before they are published. Pair repository-wide secret scanning and immediate key revocation with this narrower deployment control.
AID-D-005.007
Token, Tool-Use, Request-Parameter & Cost Spike Detection & Alerting
Medium
Profile inference request parameters and velocity at the AI gateway and provider logs to alert on max_tokens=1 credential-validation probes across multiple models as well as sudden request-rate surges (500,000+ calls in hours) characteristic of LLM Mixer pooling.
AID-E-001.001
Root & Long-Lived Credential Object Eviction
Medium
When responders identify a compromised long-lived Virtual Key, block that exact key through LiteLLM administration. Verify that each serving replica rejects the old key. Blocking it before the attacker uses it ends that opportunity. After elevation, also investigate newly issued credentials and changes to administrative access.

What Defenders Should Do Now

  • Inventory each provider's allowed identities, regions and model routes. Verify that model calls from unapproved paths fail even with a structurally valid credential. Keep billing and model enrollment permissions separate from ordinary inference.
  • Track short credential-validation bursts, abnormal token issuance, new principals and model spending together. Use enforceable quotas where supported and verify actual deny behavior. Budget alerts alone will not stop new usage. Review gateway access logs for anomalous /key/regenerate calls. If the deployment had an Enterprise license and retained LiteLLM audit logs, compare recorded key ownership before and after regeneration. Ordinary access logs may not contain those values; older audit attribution also needs corroboration because callers could override the changed-by field.
  • Contain unauthorized access at its issuer, remove confirmed attacker-created principals and runtime workloads, and separately check for persistence and unaccounted usage. For affected LiteLLM gateways, deploy a supported release containing the ownership fix introduced in v1.84.0 and restrict the key management API to independently authenticated management identities.

Conclusion

AIDEFEND  puts provider-side identity checks, enforceable usage budgets and scoped eviction at different points in the cases. Together these controls reduce the opportunity to reuse stolen credentials, continue running unapproved inference or issue fresh access through attacker-created cloud accounts.

Last updated: