Attackers have built an AI agent designed for LLMjacking, according to a report from Help Net Security. The agent takes over exposed AI models and uses them to steal credentials and mine cryptocurrency. The cost lands on the victim. For anyone running a self-hosted model or holding a paid AI API key, LLMjacking exposed AI models is now a practical risk worth understanding.

What LLMjacking Is and How the Agent Works

LLMjacking is the hijacking of a large language model (LLM) that someone else owns or pays for. Instead of buying compute or API access, an attacker finds a model that is reachable or a key that has leaked, and uses it for free. The victim pays the bill.

What is new here is automation. The reporting describes an AI agent built for this purpose. Rather than a person hunting for misconfigured services by hand, software does the finding and the taking over. The reported goals are stealing credentials and mining cryptocurrency, both of which turn a hijacked model into direct profit for the attacker.

The source describes the tool only at a high level, and we are not going to speculate beyond it. The broader point still holds: when discovery and exploitation are automated, an exposed service does not need to be interesting to be found. It only needs to be reachable.

How Exposed AI Keys and Services Get Hijacked

The suggested pattern is straightforward. An agent scans for exposed AI models and API endpoints, then takes control of what it finds. Two kinds of exposure matter most:

  • Exposed endpoints. A self-hosted model or AI interface that is reachable from the internet, often because of a convenient setting, a forwarded port, or a missing login.
  • Exposed credentials. API keys sitting in places they should not be, such as code repositories, shared config files, or old projects nobody remembers.

Once an attacker has either, the model works for them. The compute runs on your hardware or your account, and the usage shows up on your invoice or slows down your machine.

The credential-theft angle adds a second layer of concern. A hijacked AI service is not only a free resource. It can also become a way to collect more credentials, which is why one forgotten key can turn into a wider problem.

What This Means For You

If you only use AI through a consumer chat app, your direct exposure is limited. The risk grows as you do more of the following:

  • You run models at home. Home-server enthusiasts who host local LLMs, web front ends, or automation tools are the most obvious targets. An unauthenticated AI service on a forwarded port is the kind of thing an automated scanner is built to find.
  • You hold API keys. Developers and hobbyists who connect AI services to scripts, browser extensions, or side projects may have keys scattered across several places.
  • You reuse credentials. If a hijacked system exposes a password you use elsewhere, the damage spreads past the AI service. Reuse turns a single compromise into many.

The financial side is also easy to miss. Cryptocurrency mining and heavy model use consume resources quietly. You may notice only when a bill spikes, your hardware runs hot, or a provider flags unusual activity.

This fits a wider trend. Attackers are increasingly using AI offensively, not just targeting it. Our coverage of how Russian hackers used Claude AI to rebuild detected malware shows the same shift toward AI-assisted tooling on the attacker side.

How to Lock Down API Keys, Network Exposure, and Home-Server Settings

None of these steps is exotic. Most are basic hygiene applied to a newer kind of service.

Treat API keys like passwords.

  1. List every AI service you pay for or have connected to a project.
  2. Rotate keys you no longer recognize or have not used recently, and delete the ones you do not need.
  3. Keep keys out of code repositories and shared documents. Use environment variables or a password manager instead.
  4. Where your provider allows it, set spending limits and usage alerts so a hijacked key gets noticed quickly.
  5. Use a separate key for each project so one leak does not expose everything.

Reduce network exposure.

  1. Check your router for port-forwarding rules and remove any you do not remember creating.
  2. Do not expose self-hosted AI interfaces directly to the internet. If you need remote access, put it behind a VPN or another authenticated tunnel you control.
  3. Turn on authentication for every AI web interface or API you run, even if it only lives on your home network.
  4. Keep the model software and its dependencies updated.

Watch for warning signs.

  1. Unexpected charges or usage spikes on an AI account.
  2. Sustained high CPU or GPU load when you are not running anything.
  3. Logs showing requests from addresses you do not recognize.

Fix credential reuse. Use unique passwords for each service and turn on multi-factor authentication where available. If a hijacked service ever exposes a password, unique credentials keep the fallout contained.

Takeaways

The report is short, but the lesson is clear. LLMjacking exposed AI models is becoming an automated, scalable activity, and the people paying for it are the owners of the exposed keys and servers. You do not need to stop using AI tools. You do need to know where your keys live and what your network is exposing.

Today, take a few minutes to audit your AI API keys, check any self-hosted AI services you run, and review your router's port-forwarding rules. If you want more context on how attackers are putting AI to work, read our piece on AI-assisted malware development.