Safety & Security

Google Warns LLMjacking Attacks Can Burn $100,000 in Cloud Charges in a Day

Google Threat Intelligence says LLMjacking victims have lost up to $1 million, with one underground service making 500,000 API calls in under eight hours on stolen keys.

Google warns of LLMjacking spike hitting firms hard
Google warns of LLMjacking spike hitting firms hardAI-generated
By Marcus Bennett5 min read

Why it matters

  • Cloud charges from a single LLMjacking attack exceeded $100,000 in one day, per Google Threat Intelligence Group.
  • Mandiant investigated 2026 cases in which fraudulent costs reached $1 million before access was cut off.
  • A CVE-2026-40217 flaw in LiteLLM gateway software let attackers turn virtual keys into admin credentials.
  • The Jet Night service made more than 500,000 API calls in under eight hours using stolen third-party keys.
  • The report, dated November 17, 2026, describes a resale market of illicit gateways called 'LLM Mixers'.

Hijacked corporate access to large language models has produced cloud charges exceeding $100,000 in a single day, according to a Google Threat Intelligence Group report published on November 17, 2026. Mandiant, Google's incident response arm, investigated multiple cases throughout 2026 in which organisations accumulated fraudulent costs of as much as $1 million before access was cut off.

The victims span North America, South America and Asia Pacific, across financial services, retail, technology and transportation. The activity affected a range of cloud and model providers, not a single vendor. That breadth is the reason the report matters: any enterprise exposing model endpoints or mishandling cloud credentials is a potential target, and the financial damage arrives faster than in conventional cloud resource theft.

What is LLMjacking and how does it differ from cryptojacking?

The report labels the attacks LLMjacking — the hijacking of legitimate corporate access to large language models to run inference at the victim's expense. Google says the pattern differs from older resource theft such as cryptojacking because it can trigger sharp billing spikes within hours rather than gradual, concealed consumption. The ecosystem has also matured into a division of labour. One set of actors gains initial access to compromised environments; separate operators manage downstream monetisation infrastructure.

"As the LLMJacking ecosystem matures, we are also tracking a division of labor between threat actors establishing initial access to compromised environments and operators managing downstream monetization infrastructure, known as 'LLM Mixers'," said Christopher Romano, Codi Starks, Garrett Mahoney, John Wolfram, Josh Murchie, K Singh, Mathew Potaczek, Mike Lombardi and Rik Dolfing of Google Threat Intelligence Group.

How did attackers get in?

The intrusions began with exposed model endpoints and stolen credentials. The specific credential types Mandiant observed include:

  • Identity and access management (IAM) credentials — the keys that define what an identity may do inside a cloud account
  • Service account keys, used by applications rather than people
  • GitHub personal access tokens
  • Active session tokens

In several cases, attackers harvested API keys left in public code repositories and validated them with minimal test requests, deliberately keeping traffic small to avoid detection. One case began with a misconfigured service account key that granted a threat actor access to a cloud instance, followed by access to an LLM API. Another relied on a page posing as an application portal that persuaded users to paste a raw JSON payload containing an active access token — a step that let the attacker bypass both password checks and multi-factor authentication.

Elsewhere, attackers combined a stolen personal access token with the exploitation of a suspected web application flaw in a package running in a containerised cloud environment. In a separate incident, compromised corporate IAM credentials were paired with cloud federation tokens to generate fresh access keys.

What happened after the initial breach?

Lateral movement followed. Attackers searched code repositories, cloud services, databases and storage systems for additional credentials. Some queried metadata tables in cloud data warehouses looking for administrator passwords and other sensitive fields. Others cloned internal codebases and reviewed access logs for credentials tied to secondary cloud environments.

Persistence techniques were consistent across incidents:

  • Creation of unauthorised IAM and service accounts
  • Attachment of administrative policies to service roles
  • Injection of SSH keys into compute instances

One intrusion stands out for its exploitation of gateway software. A follow-on cluster abused a vulnerability in LiteLLM gateway software, tracked as CVE-2026-40217, to convert lower-level virtual keys into administrative credentials. That single step gave the attackers deep and persistent control over the compromised environment.

Monetisation came next. Attackers generated new API keys, enabled model services and ran large volumes of inference requests. One threat actor created and deleted short-lived service accounts in a "burn and rotate" approach while checking the cloud provider's billing interface to track the financial impact in real time.

What is the resale market for stolen model access?

Google Threat Intelligence Group reports signs of a broader underground market in stolen model access. The report describes "LLM Mixers" — illicit gateways that proxy, aggregate and resell unauthorised access to models. The Jet Night service illustrates the scale. According to the report, Jet Night used stolen third-party API keys to route tokens and made more than 500,000 API calls in less than eight hours. Investigators said the activity pushed close to $1 million in token costs on to the victim organisation.

The report also linked activity to open-source gateway and proxy tools, including new-api and gcli2api, which can pool stolen keys and spread requests across compromised accounts. Researchers observed custom relay infrastructure hosted on deceptive lookalike domains designed to resemble internal company hosts. Some attackers provisioned GPU-equipped instances across several regions and sought quota increases through provider APIs. In one case, a threat actor deployed custom Node.js proxy bundles and a containerised backdoor linked to the Manus agent framework inside a compromised cloud environment.

How should organisations detect and respond?

Google's guidance is blunt: defenders should treat any unauthorised enablement of LLM services as a critical security event. In several incidents, the targeted services had never been enabled before the compromise — the attackers switched them on first, then rapidly generated large bills.

The recommended controls fall into two groups. Credential and configuration hygiene:

  • Tighter credential management
  • Least-privilege access
  • Secret scanning in code repositories
  • Routine configuration audits
  • Stronger controls on API keys
  • Patching of model routing gateways and machine learning software dependencies

Detection and containment:

  • Rate limiting on model APIs
  • Network allow listing
  • Budget alerts
  • Phishing-resistant multi-factor authentication
  • Audit-log monitoring for direct model execution, reconnaissance of available models and attempts to alter logging settings

The report flagged one early-warning signal in particular: low-volume validation requests, such as single-token prompts. Attackers use them to quietly test stolen credentials before launching large-scale abuse. A single-token request that no legitimate workload would send can precede a six-figure invoice by hours.

For enterprises, the economics are the stakes. Inference costs scale with request volume, and an attacker holding a valid key faces no natural ceiling. Google's report implies that as long as stolen LLM access can be aggregated and resold through LLM Mixers, the underground market will keep rewarding credential thieves — and organisations that have not enabled LLM services may discover their first model usage on the invoice.

Source: itbrief.com.au

Share this article:

More from Marcus Bennett

Marcus Bennett

Show full bio

Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

718 articles

Related articles

  1. Stolen AI Credentials Fuel a Booming Black Market in LLMjacking
  2. Stolen Enterprise AI Access Sells for Ten Cents on Underground 'LLM Mixers'
  3. Google Cloud warns attackers will turn from data theft to AI decision manipulation
  4. Fireworks AI Says Attackers Reached Datasets and Model Weights
  5. Microsoft Disrupts EvilTokens, AI Chatbot Platform Behind 12,000 Account Hacks

« Previous articleNext article »