Disclosure: this guide summarises research from Google Threat Intelligence Group and Mandiant, CrowdStrike, Sysdig and the Cloud Security Alliance as reported by security outlets. We have not investigated these incidents ourselves. Figures are attributed to their sources.
LLMjacking attacks are what happens when someone steals your cloud or AI credentials and uses them to run language models on your account. In a report published at the start of November 2026, Google Threat Intelligence Group and Mandiant said some organisations were losing more than $100,000 a day to these attacks, with total fraudulent charges approaching $1 million in individual cases.
The honest summary: this is mainly a billing crime, not a data breach, and that is exactly why it slips past many security teams. Every request looks legitimate. The first sign is usually the invoice. The good news is that the most effective defences are cheap and boring: spending caps, narrowly scoped keys, short-lived credentials and alerts on unusual usage. Most victims had none of them in place.
Table of Contents
- What LLMjacking Attacks Are
- What Google Found About LLMjacking Attacks
- How LLMjacking Attacks Work, Step by Step
- The Resale Market Behind LLMjacking Attacks
- Why LLMjacking Attacks Are Hard to Spot
- LLMjacking Attacks Become Attack Infrastructure
- Eight Defences Against LLMjacking Attacks
- Signals of LLMjacking Attacks to Monitor
- Responding to LLMjacking Attacks
- An AI Key Policy That Stops LLMjacking Attacks
- LLMjacking Attacks on Small Teams
- FAQ
- Verdict
- Sources and Further Reading

What LLMjacking Attacks Are
The term LLMjacking was coined by Sysdig’s threat research team in May 2024, according to the Cloud Security Alliance, to describe a then-new attack: compromising cloud credentials or API keys specifically to consume AI inference. Two years later it has its own tools, its own underground market and its own threat actor designations.
The mechanics of LLMjacking attacks are simple. AI models are expensive to run, and access is sold by the token. If an attacker gets hold of a key that can call a model, they get free use of something valuable, and the victim pays. The Rankiteo write-up of one AWS case put it well: unlike cryptomining or data theft, LLMjacking abuses the victim’s billing relationship with the cloud provider, turning unauthorised AI usage into a revenue stream.
CrowdStrike’s 2026 Threat Hunting Report describes a closely related technique it calls cost harvesting. Both exploit access to corporate AI infrastructure; the difference is mostly emphasis, whether the goal is using the compute or simply draining the account.
What Google Found About LLMjacking Attacks
The GTIG report, titled LLMJacking: Monetizing Compromised Cloud Models, covers incidents Mandiant responded to during 2026 across several cloud and model providers. Help Net Security reported that victims spanned financial services, retail, technology and transportation, in North America, South America and Asia Pacific.
The headline case involved a service called Jet Night, which GTIG described as a legitimate token-routing service being fed with stolen third-party API keys. Using those keys, it executed more than 500,000 model calls in under eight hours, leaving the victim with nearly a million dollars in token costs.
The report also tracked a threat actor it designates UNC6993. According to SiliconANGLE and CSO Online, the group used tools including new-api and gcli2api, likely to pool stolen keys and spread requests across many accounts, and GTIG believes it sells access through underground storefronts. When it got into cloud projects, it tried to create additional API keys and switch on AI services that had never been enabled.
One incident involved a software vulnerability. CRN Asia and SiliconANGLE reported that attackers exploited CVE-2026-40217 in an AI gateway’s key regeneration endpoint to turn ordinary virtual keys into administrator credentials for the victim’s gateway. If you run an AI gateway, check whether you are on a patched version.
How LLMjacking Attacks Work, Step by Step
1. Get a credential
GTIG found the most common starting point was exposed or stolen cloud credentials: IAM credentials, service account keys, GitHub personal access tokens and active session tokens, as well as AI provider API keys directly. Keys leak from public code repositories, misconfigured storage buckets, container images, IDE plugins and browser extensions.
2. Validate it
Attackers test stolen keys against a list of providers. The Cloud Security Alliance described campaigns validating keys against OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Cloud Vertex AI, Mistral and others. Validation often uses tiny requests, sometimes a single token, which is one of the signals defenders can watch for.
3. Expand access
Once inside a cloud account, attackers create their own keys and enable AI services the victim never used. Reworked highlighted the consequence: deleting the original leaked key does not help if the thief has already created fresh ones.
4. Evade limits
To get around rate limits, GTIG observed attackers repeatedly creating and deleting temporary service accounts. In AWS cases, attackers generated Bedrock service-specific API keys as an additional, less detectable way to invoke models.
5. Monetise
Access is either used directly or resold through proxies, sometimes called LLM mixers, that relay paying customers’ prompts through the stolen accounts.
Speed is part of what makes LLMjacking attacks dangerous. A September report from Cybersecurity News described exposed keys being hijacked within 19 minutes. If a key reaches a public repository, assume it is compromised immediately, not when you notice.
The Resale Market Behind LLMjacking Attacks
The economics explain why LLMjacking attacks are growing. A Hugging Face community post reported stolen LLM API keys selling for as little as $30. GTIG said Jet Night sells an unlimited plan claiming frontier-model access for about $60 a month, against $100 to $250 for comparable consumer plans, and concluded that prices that low only work on stolen keys.
BleepingComputer described the role of LLM mixers: services that offer access to AI resources at very low prices by relaying customers’ prompts to hijacked accounts. The buyer may not know or care where the access comes from. The victim pays the difference.
Decryption Digest reported operations simultaneously targeting more than 30 AI providers. LLMjacking attacks are an industry, not a handful of opportunists.
Why LLMjacking Attacks Are Hard to Spot
Business Standard’s summary of the CrowdStrike findings made the central point: the individual actions look legitimate. An API request to a model is not suspicious in itself. Neither is a cloud login, or a request with a valid token. There is no malware to detect and often no data leaving the network.
That puts the burden on usage and billing monitoring rather than traditional intrusion detection. Many organisations do not watch AI spend closely, and many have no budget alerts at all. By the time someone in finance asks why the cloud bill tripled, the attackers have had weeks.
The spread of AI keys makes it worse. Developers create keys for experiments, prototypes and personal tools, often with broad permissions and no expiry. Every one of those is a potential entry point.
LLMjacking Attacks Become Attack Infrastructure
The Cloud Security Alliance flagged a worrying shift in June 2026. A Sysdig finding documented stolen AI compute being used not just as resold capacity but as the reasoning engine wired into an autonomous attack pipeline. In other words, attackers used hijacked models to power their own offensive tools.
That raises the stakes of LLMjacking attacks beyond the bill. A hijacked account may be generating phishing content, malware or reconnaissance in your name, on your infrastructure. Auth0’s analysis made a related point: a compromised AI credential can give an attacker far more than raw compute, depending on what the model can access through tools and connected data.
For organisations running agents with tool access, a stolen key is not just a way to run a model. It may be a way to use your agent’s permissions.
Eight Defences Against LLMjacking Attacks
- Set hard spending caps. GTIG lists spending caps among its recommendations. A cap turns a six-figure loss into a capped one and an alert.
- Scope keys narrowly. Limit each key to specific models and environments, as GTIG recommends. A key for a staging chatbot should not be able to call every model in production.
- Prefer short-lived credentials. Workload identity and temporary tokens beat long-lived static keys wherever your platform supports them.
- Disable unused AI services. Attackers enable services you never use. Block them at the organisation or project level.
- Scan for leaked secrets. Use secret scanning on repositories, container images and CI logs, and revoke anything found immediately.
- Restrict who can create keys. Many incidents involved attackers minting new keys. Limit that permission tightly and alert on every new key.
- Patch AI gateways. The CVE-2026-40217 case shows gateway software is part of the attack surface. Keep it current.
- Centralise AI usage. Route model calls through one gateway or account where usage is visible, rather than scattering keys across teams.
Signals of LLMjacking Attacks to Monitor
GTIG specifically recommends log monitoring for the single-token pings attackers use to validate stolen keys. Beyond that, the useful signals are about change:
- A sudden jump in token usage or requests per hour.
- Calls to models or regions your applications never use.
- New API keys or service accounts you did not create.
- AI services being enabled in projects that did not have them.
- Requests arriving at hours when your applications are idle.
- Many rapid creations and deletions of service accounts.
Most cloud providers offer budget alerts and anomaly detection on billing. Turning those on is the single cheapest step most organisations can take today.
Responding to LLMjacking Attacks
Act in this order. First, revoke the compromised credential and every key created from that account, not just the one you found. Second, check for new service accounts, keys and enabled services, and remove anything you did not create. Third, apply a spending cap so the damage stops growing while you investigate.
Then contact your cloud and AI providers. Many have processes for fraudulent usage, and some will review charges caused by confirmed credential theft, though there is no guarantee. Document the timeline and the evidence before you clean up, because you will need it.
Finally, find out how the credential leaked. If it came from a public repository, assume other secrets in the same repository are also compromised and rotate them too.
An AI Key Policy That Stops LLMjacking Attacks
Most LLMjacking attacks in these reports trace back to ordinary habits: a key pasted into a script, a token committed to a repository, a personal experiment with production credentials. A short written policy prevents most of them.
Keep it to a page. Say where keys may be stored, which is a secrets manager, never code or chat. Say who may create keys and for what. Require an expiry date on every key, even long ones. Require a named owner, so that when a key shows up in a scan, someone is responsible for it.
Separate environments strictly. Development, testing and production should use different accounts or projects with different keys and different budgets. A leaked development key should never be able to spend production money.
Make spending visible. A weekly summary of AI spend by team or project, shared openly, catches anomalies faster than any tool, because people notice when their own number jumps.
Review quarterly. List all active keys, delete the ones no longer used, and check that every remaining key still has an owner. It takes an hour, and it removes exactly the forgotten credentials that attackers find.
Explain why. Developers follow security rules more readily when they understand the risk. Sharing a real example, such as half a million calls in eight hours on someone else’s key, makes the case better than any policy wording.
LLMjacking Attacks on Small Teams and Solo Developers
The large cases make headlines, but LLMjacking attacks hit individuals and small teams too, and the damage can be proportionally worse. A freelancer who leaks a personal API key in a public project has no security team watching the logs and may have a card on file with no limit.
The defences scale down well. Use a prepaid balance or a low monthly cap on personal AI accounts, so a stolen key can only spend what you loaded. Keep keys in environment variables or a password manager, never in code. Add a pre-commit hook or a free secret scanner to your repositories. And create separate keys for each project, so you can revoke one without breaking everything else.
If you publish code examples, double-check them before pushing. Many leaked keys come from tutorials and demo repositories where someone pasted a real key to get a quick example working and forgot to remove it.
Finally, check your provider dashboards once a week. LLMjacking attacks are easy to miss when nobody looks, and easy to catch when someone glances at a usage graph and sees a spike they did not cause.
Frequently Asked Questions About LLMjacking Attacks
What are LLMjacking attacks?
Attacks in which criminals steal cloud or AI credentials and use them to run language models on the victim’s account, either for their own use or to resell access. The victim pays the bill.
How much can they cost?
Google Threat Intelligence Group reported some organisations losing more than $100,000 a day, and one case with nearly $1 million in token costs from over 500,000 calls in under eight hours.
How do attackers get the keys?
Usually from exposed or stolen cloud credentials, service account keys, GitHub tokens, session tokens and AI API keys leaked in code, storage buckets, containers or plugins.
Why are they hard to detect?
Each request looks legitimate, and there is often no malware or data theft. Detection depends on watching usage and billing for unusual patterns.
What is the best single defence?
A hard spending cap with alerts. It does not prevent the theft, but it limits the loss and tells you quickly that something is wrong.
Who is behind them?
GTIG tracks a group it calls UNC6993 among others, and describes an underground market where stolen access is pooled and resold cheaply through proxy services.
Verdict
Treat LLMjacking attacks as a billing risk that needs the same attention as any other fraud. Set spending caps today, scope every key to the models and environments it needs, alert on new keys and unusual usage, and remove long-lived credentials wherever you can.
The honest caveat is that the threat is moving. Stolen AI compute is now being used to power attack tools, not just resold, and new vulnerabilities in AI gateways keep appearing. The figures here come from incident response firms and represent cases they handled, which skews toward large losses.
Next step: this afternoon, turn on budget alerts and a hard cap on every AI and cloud account you control, then run a secret scan across your repositories and revoke anything it finds.
Sources and Further Reading
- BleepingComputer: $1 million in AI charges, for the GTIG findings and LLM mixers.
- SiliconANGLE: $100K a day, for UNC6993, Jet Night and recommendations.
- CSO Online: AI they never ordered, for initial access methods.
- Cloud Security Alliance: LLMjacking evolved, for history and the attack-pipeline shift.
- Auth0: the hidden cost of a stolen key, for impact beyond the bill.
On this site: learn from the Fireworks AI breach, cap spend with LLM API cost optimization, secure agents with prompt injection and agent security, and read about the ARTEX AI FBI warning.

