LLMJacking is a form of cybercrime in which attackers steal access to paid artificial intelligence systems and make someone else cover the cost. A criminal who obtains an organisation’s cloud credentials or AI service keys may be able to run large volumes of model requests without paying the provider directly. Google’s threat investigators reported cases in which that activity generated substantial unexpected bills, while illicit intermediaries offered stolen access to other buyers.
This is a different kind of loss from a traditional data breach. The victim may still have its files and its service may appear to work, yet a stranger is quietly using its paid capacity. Google describes incidents affecting companies across several industries and regions in its November 2026 investigation into stolen AI access. The company reported individual cases involving fraudulent usage charges exceeding $100,000 in a day. That figure describes observed cases, not the typical bill for every affected customer.
LLMJacking turns an API key into a cash machine
An application programming interface, or API, lets software send requests to another service. A business might use one to summarise customer messages, search internal documents or power an automated support desk. The organisation usually pays for the computing work consumed by those requests. A secret key identifies the account that is allowed to use the service.
If that key is exposed in a code repository, a poorly protected server or a compromised account, an attacker may be able to impersonate the authorised customer. The provider then sees apparently valid requests, even if the true customer did not make them. The criminal does not need to steal physical servers to turn another organisation’s credit into an illicit service.
This is not a claim that every chatbot account can be monetised in the same way. The exact exposure depends on the service, subscription, permissions and billing arrangements. An ordinary consumer logging into a chat app has a different risk profile from a company whose cloud account can call powerful paid models automatically.
Why the stolen access is valuable to criminals
A language model consumes computing resources each time it processes input and produces output. For businesses, those costs can be small or large depending on the application and contract. A stolen account with generous limits offers a way for someone else to run work without paying the full cost themselves.
Google’s investigators describe middlemen called LLM Mixers. Such services collect or proxy access from compromised accounts and route requests for customers of the illicit service. The model provider may see traffic arriving under a real company’s credentials, while the people buying the underground service are hidden behind another layer.
This division of labour is important. Some criminals specialise in stealing or buying credentials, while others handle the resale infrastructure. A business can therefore face abuse from people it never directly encountered. Removing one visible account may not stop the underlying market if other credentials remain available to the same operation.
How investigators see the attack developing
According to Google’s research, attackers have compromised cloud resources and model-access arrangements in multiple sectors, including finance, technology and retail. Security teams were able to study misuse patterns involving abrupt increases in model calls and unusual use of stolen credentials. Google says the fraudulent charges in some cases mounted rapidly, before the organisation removed access.
In a later Google threat-tracker report, the company described how LLM Mixers distribute unauthorised requests across pools of stolen credentials and computing resources. That can complicate detection because a suspicious user’s work may be spread among victims rather than concentrated on one account. The report is threat-intelligence evidence from a technology vendor, not a count of every such attack worldwide.
The technical label can make the crime sound exotic. In ordinary terms, it is similar to discovering that someone has stolen a company credit card and is repeatedly charging a costly service to it. The difference is that the resource being purchased is computing capacity, and the charges may arrive through a cloud billing account rather than a shop receipt.
The first warning may be a cloud bill
Some victims might notice the problem through unexpectedly high usage rather than a security alert. A workload that usually sends modest traffic can suddenly show sustained calls at unusual hours. Logs may reveal models or regions the business never meant to use. The warning signs are specific to the architecture, so no single threshold should be advertised as a universal detector.
A large bill should prompt more than a dispute over money. The business needs to establish which credentials were used, what those credentials could access and whether the same attacker may have reached data or other services. The invoices, system logs and authentication trail answer different questions, and all are worth preserving for the investigation.
It is also possible for a genuine application bug to create high usage. Repeated requests, poorly controlled loops and unexpected customer demand can produce charges without any criminal intrusion. Security teams should investigate before concluding that every cost spike is evidence of a hacker.
What companies can do before and after exposure
The first line of defence is limiting the power of credentials. An AI key intended for one application should not automatically have permission to manage every cloud resource or view every dataset. Storing secrets in approved systems, rotating them when compromised and restricting where they can be used reduces the damage from any single mistake.
Usage monitoring is a financial control as well as a cybersecurity one. Organisations can define sensible spending alerts and examine unusual increases quickly. Where supported, quotas, account separation and provider-side restrictions can make a stolen key less useful. A small company using only one AI feature rarely needs an unrestricted key that can trigger unlimited costs.
LiveAIWire’s coverage of autonomous AI agents involved in cybersecurity incidents explored the wider question of what happens when software obtains authority to act. LLMJacking highlights the opposite problem: an outside criminal obtains the authority that a legitimate AI application was supposed to have. In both cases, permissions and monitoring matter more than whether an interface looks intelligent.
Another useful comparison is the insider-risk warning around Microsoft AI agents. Organisations need to know where identities, automation and sensitive material meet. Security design that assumes every request authenticated by a valid key is benign is weaker than design that also evaluates the purpose, scale and context of usage.
What a business should do when it discovers misuse
Once suspicious activity is confirmed, the organisation should revoke or disable affected credentials through its provider, investigate adjacent accounts and preserve evidence. It may need to rebuild exposed applications, remove a secret from public repositories and check whether attackers planted additional access mechanisms. Removing the visible key alone is not enough if the original route of compromise remains open.
Billing teams should contact the provider promptly with dates, usage records and the steps already taken. Any request to reverse charges depends on the contract and the provider’s assessment; reimbursement should never be assumed. Customers or regulators may also need notification if the investigation uncovers personal-data exposure, but excessive usage by itself does not prove such exposure.
A related strand of LiveAIWire reporting, the warning about emergency controls for autonomous agents, is relevant to incident response. Automation needs a reliable way to stop, and staff must know who is authorised to use it. That principle extends to model subscriptions, secrets and cloud resources that could otherwise keep spending after a compromise begins.
The lesson is about ordinary controls around extraordinary technology
AI can make cloud infrastructure feel novel, but this attack rests on familiar weaknesses: stolen credentials, excessive permissions, poor monitoring and profitable resale. New models make the underlying resource especially valuable, and criminals adjust to that value. They do not need to create a new form of intelligence to exploit the bill.
For users of small-business AI tools, the useful question is not whether every service is dangerous. It is who controls the account, what can be charged to it and how quickly unusual activity would be noticed. A business that knows those answers can treat LLMJacking as a manageable security exposure rather than an inexplicable surprise at the end of the month. The same review should extend to contractors and external developers who deploy integrations using the company’s accounts. A key issued for a short project may remain active after the work ends unless someone takes responsibility for closing it. Regular access reviews can remove that quiet risk before criminals find it.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.
