Skip to content
AI SecurityLLMllm-securitycredential-theftthreat-intelligencecritical

LLMjacking and LLM Mixers: Mandiant finds $1M average victim cost

3 min read
Share

Google Threat Intelligence Group and Mandiant published a joint assessment on LLMjacking on November 17, 2026. The report is the most detailed quantification of the attack category to date. The average victim incurs between $800,000 and $1.2 million in cloud compute charges before the unauthorized usage is detected and stopped.

What LLMjacking is

LLMjacking is the unauthorized use of cloud-hosted large language model APIs using stolen credentials. Attackers compromise AWS, Azure, or GCP accounts that have permission to invoke LLM services such as Amazon Bedrock, Azure OpenAI, or Vertex AI, then use those credentials to run inference at the victim's expense. The victim is billed for the compute while the attacker uses the model for whatever purpose they choose, from reselling access to running their own AI-powered services.

UNC6993 and the LLM Mixers marketplace

Mandiant attributes a significant portion of recent LLMjacking activity to a threat cluster tracked as UNC6993. This group operates a marketplace called LLM Mixers on underground forums. LLM Mixers functions as a reseller of stolen inference capacity: sellers list cloud API credentials that have been validated against specific LLM API endpoints, with each listing specifying the model name, available context window size, and estimated remaining budget. Buyers pay in cryptocurrency and receive credentials with which they can interact directly with the LLM service.

Why the costs are so high

LLM API calls are significantly more expensive per unit than general compute resources. A single high-token inference request against a frontier model can cost more than a full hour of a large EC2 instance. Attackers often run workloads that maximize token throughput, and the absence of spend-specific alerts on AI services means victims often do not notice until their monthly billing cycle. By then the cumulative charges can reach seven figures.

Detection and prevention

Mandiant released a public YARA rule for LLM Mixers tooling with rule ID UNC6993-LM-001. For prevention, audit all IAM roles and service accounts with permissions to invoke Amazon Bedrock, Azure OpenAI, or Vertex AI endpoints. Rotate API keys that grant these permissions on a regular schedule. Enable spend anomaly alerts on all AI services; both AWS and Azure provide budget alerting that can trigger at a percentage threshold before the billing period closes.

Durable lesson

Credential hygiene for AI services requires the same rigor as credential hygiene for compute resources. The access control surface expanded when organizations began onboarding LLM services, and in many cases the IAM policies were written quickly during proofs of concept and never tightened for production. This report is an opportunity to revisit those permissions before an attacker finds them first.

Related articles