LLMjacking: When a Leaked AI Key Becomes a $1 Million Bill

Google reports attackers hijacking company AI access, with bills past $100,000 a day. What it means for Indian teams using LLM APIs, and what to check first.

Google's threat intelligence team says some of the companies it studied were charged more than USD 100,000 in a single day for AI usage they never made. At a few organisations the total reached USD 1 million before anyone cut off access.

If your business pays for a large language model (LLM) through an API, that sentence is worth reading slowly. The model is not the weak point. The key that lets your software talk to it is.

Why a billing line became a security problem

The Hacker News and IT Brief Asia both report on a Google Threat Intelligence Group and Mandiant analysis of 2026 cases. The attack is called LLMjacking. An attacker gets hold of a company's API key or cloud login, then runs their own AI workloads on it. The company pays.

The cases covered North America, South America and Asia Pacific, across financial services, retail, technology and transportation. The sources do not break the numbers down for India, so we will not guess. But Asia Pacific is in the list, and the mechanics do not depend on geography.

Think of a shop that gives a delivery boy a card for the fuel pump. If the card is copied, the shop owner sees the bill at month end. LLMjacking is the same arrangement, except the pump runs at machine speed. Google says that, unlike cryptojacking, where a hijacked server mines coins slowly, charges here can spike within hours.

How the keys get out

None of the reported entry points were exotic. Google lists phishing, leaked source code, misconfigured cloud settings and exposed client-side code. Some keys were simply sitting in public code repositories, and attackers tested them with tiny requests so nothing looked unusual.

Two cases are worth knowing by shape. In one, a fake application portal asked users to paste a block of text that contained a live access token. That skipped the password and the multi-factor check entirely, because the token already proved the login had happened. In another, a stolen GitHub personal access token opened private repositories, and the attackers then went looking inside the code for more secrets.

Once inside, the reported pattern was to search code, databases and storage for further credentials, create new service accounts, and in some cases attach administrative rights to them. One intrusion used a flaw, tracked as CVE-2026-40217, in LiteLLM, a gateway that routes requests to different models. It let low-level virtual keys be turned into administrator keys. One actor used a "burn and rotate" approach: create a short-lived service account, use it, delete it, repeat, while watching the provider's billing page.

The resale market behind the bills

The part that changes the economics is the second group of criminals. Google describes "LLM Mixers", services that proxy, aggregate and resell stolen model access. One of them, Jet Night, reportedly advertised an "unlimited" plan for about USD 60 a month.

IT Brief Asia reports that Jet Night used stolen third-party keys to make more than 500,000 API calls in under eight hours, pushing close to USD 1 million in token costs onto one victim. Open-source tools named in the report can pool stolen keys and spread requests across many compromised accounts.

So the person who steals your key may never use it. They sell it. That makes even a small company a worthwhile target, because the buyer only needs a cheap pass to a frontier model.

What we do not know yet

The reports do not say how many organisations were hit, how often the money was recovered, or how cloud providers handled the charges afterwards. They do not say whether small Indian businesses were among the victims. The cases were investigated by Mandiant, so they skew towards organisations that call in a major incident team.

Google also expects attackers to prefer LLMjacking over older resource theft as companies adopt higher-tier models and agent-style workloads. That is a forecast, not a measurement.

Three things to try this week

Google's own advice reads like a checklist, and parts of it can be tested in an afternoon.

  1. Find your keys. Write down every place an LLM API key lives: servers, laptops, shared documents, chat threads. Then search your code repositories for them. Public ones were a real source of leaks in these cases.
  2. Set a budget alert you would notice. Google recommends budget alerts, rate limits and network allow lists. Pick a daily spend figure that would surprise you, and make sure the alert reaches a phone, not a mailbox nobody opens.
  3. Ask who enabled what. In several incidents the AI services had not been switched on before the break-in, so Google advises treating any unexpected enabling of an LLM service as a critical event. Check which AI services are enabled in each cloud project, and who has the right to enable more.

If you run a gateway such as LiteLLM, add its patch status to the same list.

Do you know today, without looking, what your AI spend was yesterday?

By Nikhil Teltumbade, Director, Godrix AI Solutions.
Working on something similar? Godrix AI Solutions builds AI automation and apps for business workflows. Book a free audit.

← All posts