LLM Jacking: The New Threat Targeting AI Systems

AI has become one of the most valuable technologies for businesses. But there is another side to that value.

Hackers are starting to treat AI accounts, models and computing power as assets worth stealing.

Instead of spending hundreds of dollars on premium AI tools or massive cloud computing resources, attackers can simply look for someone else's access.

This growing threat is known as LLM-jacking.

And it is creating a new underground market where stolen AI access is being bought, sold and used for everything from cybercrime to espionage.

What Is LLM Jacking?

LLM-jacking is the unauthorised use of someone else's large language model (LLM) accounts or computing resources.

Attackers can steal credentials for AI platforms or compromise servers and cloud environments that have access to powerful AI models.

The goal is simple: use expensive AI resources without paying for them.

Security researchers have reported a significant increase in this activity, pointing to the emergence of an underground economy built around stolen AI access.

In some cases, attackers don't even need to steal an AI account. They can compromise a company's infrastructure and use its computing power to run their own AI models.

Either way, the victim pays the bill.

How Does LLM Jacking Work?

There are several ways attackers can get their hands on AI resources.

One approach is stealing login credentials through phishing, malware, infostealers or other credential-theft techniques. Once attackers gain access, they can use the victim's subscription or API resources.

Another approach targets the infrastructure itself.

Attackers may break into cloud-hosted servers, deploy AI models and quietly consume the organisation's computing resources.

It is similar to cryptojacking, where compromised computers are secretly used to mine cryptocurrency.

The difference?

Instead of mining coins, attackers are now putting stolen computing power to work for AI.

Stolen AI Access Is Becoming a Business

Perhaps the most concerning part is that stolen AI access is no longer simply being used privately.

It is being sold as a service.

Google Threat Intelligence researchers have reported underground marketplaces offering access to AI services from major providers at discounts reportedly reaching as high as 97%.

Some sellers even advertise replacement credentials if an account gets blocked.

That means attackers can potentially access expensive AI capabilities for a fraction of the legitimate cost.

For cybercriminals, the economics are difficult to ignore.

Why Are Hackers Targeting Company Servers?

AI models can require significant computing resources to operate.

For attackers, obtaining that infrastructure legitimately can be expensive.

Compromising someone else's infrastructure changes the equation.

The attacker gets the computing power.

The company gets the cloud bill.

This makes organisations running AI workloads an increasingly attractive target.

And there is another problem: AI workloads can look legitimate.

A company may already expect high CPU, GPU, memory and network usage because its AI systems are running normally.

An attacker hiding inside that activity may therefore have a much easier time avoiding attention.

What Are Criminals Using AI For?

AI is already being explored and misused by different threat actors.

Attackers can potentially use AI tools to assist with activities such as:

• Generating phishing and social-engineering content

• Writing or modifying malicious code

• Automating research and reconnaissance

• Creating convincing scam communications

• Processing stolen information

• Supporting cybercrime operations at scale

The availability of stolen AI access can make these capabilities cheaper and easier to obtain.

That creates an uncomfortable imbalance: attackers may use stolen technology to attack organisations that are spending money to defend themselves against it.

Why Is LLM Jacking Difficult to Detect?

The biggest challenge may not be identifying that an AI system is being used.

It is identifying who is using it and why.

As organisations deploy more AI models internally, increased computing activity becomes normal.

A sudden spike in GPU usage might be legitimate.

A new model being deployed might be legitimate.

Large amounts of data moving through an AI environment might also be legitimate.

Attackers can exploit this noise.

If malicious activity looks like normal AI activity, traditional monitoring may not immediately raise an alarm.

What Should Companies Watch For?

AI security can no longer stop at protecting passwords and API keys.

Organisations should also monitor the infrastructure behind their AI systems.

Key warning signs include:

• Unusual AI account or API activity

• Unexpected increases in cloud or GPU usage

• New or unauthorised AI models appearing on servers

• Unknown accounts accessing AI infrastructure

• Unusual data transfers or workloads

• Sudden changes in computing patterns

• Credentials being used from unexpected locations or environments

Most importantly, organisations need visibility across AI accounts, identities, applications, cloud infrastructure and computing resources.

AI Is Now an Asset. Treat It Like One.

The AI arms race isn't happening only between businesses.

Cybercriminals are finding ways to exploit the same technology and infrastructure businesses are investing in.

LLM jacking is a warning that AI resources themselves can become targets.

Protecting AI therefore means more than securing the model.

It means knowing who has access, what they are doing, where the workload is running and whether the computing activity makes sense.

Because when AI becomes valuable enough to steal, every AI workload becomes part of your attack surface.

When AI becomes a target, having the right incident response team matters. WhiteKnight helps organisations investigate, contain and respond to cyber incidents before they escalate.

And with attackers constantly adapting their tactics around AI, understanding AI Speed Gap: Attackers Adapt Faster Than Defenders 2026 becomes an important part of seeing where the threat landscape is heading.