LLMJacking

LLMJacking

LLMJacking refers to an attack method in which criminals use stolen credentials to run powerful AI language models at someone else's expense. The resulting compute costs can climb into the tens of thousands of dollars within just a few days.

Large language models — meaning AI programs that can understand and generate text — are expensive to operate. Anyone who wants to use them books access with a provider and pays per request. LLMJacking refers to the attack in which criminals steal someone else’s credentials to these services and then use them themselves — billed to the victim. The name is modeled on the term carjacking, i.e. stealing a car with a key rather than a crowbar: the account isn’t hacked, it’s effectively hijacked. The attackers act as if they were the rightful user.

Why stolen AI access is so valuable

A single compromised API key — a digital passcode that lets software access an AI service — can generate costs of several thousand dollars per day. Security researchers at the company Sysdig observed cases in 2024 where the potential damage exceeded 46,000 dollars daily. The victim often only notices the misuse when the monthly bill arrives.

The appeal for attackers is considerable: they gain access to computing power they don’t have to pay for themselves. They use this power either to sell access on to other criminals, or to run their own services with it — for example, to write malware or automate fraud schemes.

How an LLMJacking attack unfolds

The attack almost always begins with the theft of an API key. Such keys are often left unprotected in publicly accessible code repositories — online archives where developers store their code. Sometimes they also end up in the wrong hands through other data leaks. Automated search scripts scour the internet around the clock looking for such keys.

Once a key is found, the attackers automatically check which services it can reach and which models are enabled. They then put their own interface in front of it — a so-called proxy — and resell the access to third parties. The entire system runs without the actual account holder noticing anything.

LLMJacking differs here from classic data theft. The attackers don’t want personal information from the account, only the computing power attached to it. This makes the attack unusual, but also harder to detect: nothing is stolen that goes missing — something is used that remains in place.

Where LLMJacking shows up in practice

Those affected are mainly developers working with AI services from major providers — such as OpenAI, Anthropic, or Google. These services issue API keys that let developers connect their own applications to the AI models. Anyone who carelessly stores such a key in a public GitHub repository risks becoming a victim within minutes. Automated scanners search such platforms nearly in real time.

The term LLMJacking first appeared in trade publications in 2024, when security researchers documented organized campaigns specifically targeting cloud accounts with AI permissions. Since then, several major providers have introduced automatic detection systems that flag unusually high usage. Nevertheless, the attack surface remains large, because more and more applications are integrating AI models.

For companies, LLMJacking is therefore not just a technical problem but a financial one as well. Unlike with many other attacks, the damage doesn’t arise from data loss, but from a bill waiting at the end of the month.

Related Products

Latest News

Subscribe free. Unsubscribe the second it sucks.

High-signal news across AI, business, UX, and tech. Every morning.