Threat actors are systematically infiltrating enterprise artificial intelligence infrastructure to hijack compute access, generating massive financial losses for victim organizations. In several recent intrusion cases investigated by cybersecurity firm Mandiant, unauthorized access to corporate AI models caused fraudulent cloud bills to exceed $100,000 in a single day, with cumulative charges reaching as high as $1 million, as reported by SecurityWeek.

The activity, commonly referred to as LLMJacking, represents an evolution of resource-theft tactics previously dominated by cryptojacking. Instead of diverting cloud virtual machines toward cryptocurrency mining, adversaries are targeting large language model allocations to siphon expensive API tokens and high-performance compute capacity.

Initial Access and Privilege Escalation Tactics

According to Mandiant's findings, the primary initial access vector in these campaigns stems from exposed or compromised cloud credentials and leaked API keys. Attackers actively sweep public code repositories, exposed systems, and cloud environments to locate lingering secrets tied to frontier model subscriptions.

Once initial footholds are established, attackers have demonstrated the ability to escalate privileges within intermediate AI tooling. In one documented intrusion, threat actors exploited a vulnerability in LiteLLM, cataloged as CVE-2026-40217. By exploiting this flaw, the operators upgraded standard, low-privilege API keys into full administrative credentials, giving them broad control over the target organization's AI deployments.

The Rise of Underground LLM Mixers

Stolen AI access is not solely consumed by the initial intruders; it is also fueling a burgeoning shadow economy. Researchers at the Google Threat Intelligence Group discovered that compromised enterprise credentials are increasingly routed into commercial underground services known as LLM Mixers.

One such operation identified by researchers is Jet Night, a token routing service that advertises unlimited subscription plans offering purported access to frontier language models for approximately $60 per month. Analysts believe services like Jet Night rely directly on stolen enterprise tokens to fulfill customer queries while offloading the substantial inference costs onto breached corporate accounts.

What You Should Do

Organizations integrating large language models into internal systems and consumer-facing applications should implement immediate defensive measures to detect and prevent unauthorized API use.

  • Patch LiteLLM deployments: Ensure all instances of LiteLLM are updated immediately to remediate the CVE-2026-40217 privilege escalation vulnerability.
  • Audit secrets and cloud credentials: Scan internal and external source code repositories, continuous integration pipelines, and configuration files for exposed AI service tokens and cloud management keys.
  • Enforce strict least privilege: Restrict AI API keys to specific endpoints and operations rather than deploying unrestricted or administrative tokens across standard workloads.
  • Configure real-time billing alarms: Set aggressive spend caps and real-time alerts within cloud and AI vendor management consoles to detect unexpected consumption spikes before costs escalate into five- or six-figure territory.
  • Monitor API consumption anomalies: Establish baseline usage profiles for corporate AI endpoints and inspect unusual traffic patterns, including sudden surges in token consumption or requests originating from unexpected IP ranges.