Google Threat Intelligence Group and Mandiant published research into LLMJacking early November 2026. Criminals are using stolen cloud credentials or API keys to run AI models on a company’s account, gaining access to large language models (LLMs) and expensive GPU processing capacity while the company foots the bill. Google found fraudulent charges in individual accounts exceeding $100,000 in a single day, with cumulative charges at some companies reaching $1 million before access was removed. The model resembles cryptojacking but abuses paid model resources and can drive rapid increases in cloud costs within hours.
Across Mandiant’s LLMJacking investigations, the company found that the most common entry route was through exposed or stolen cloud identity tokens. These include service account keys, GitHub personal access tokens and active session tokens. One group harvested corporate API keys from public GitHub repositories. The keys typically leak when code is pushed to a public repository without being checked for sensitive information or when session tokens are stolen by malware on a device, or when credentials from earlier breaches are traded on the dark web. Each route gives the attacker valid credentials, so they are simply logging in, and these compromised credentials gave attackers an initial foothold which they then used to expand their permissions, create additional keys or enable LLM services.
In Kenya, credential-related attacks are on the increase. The National KE-CIRT/CC detected 44,949,518 brute-force attack attempts between July and September 2026, an increase of 85.41% from April to June. The report points out that these attacks primarily targeted cloud service providers and government systems, identifying compromised credentials, a lack of multifactor authentication and insecure remote access configurations among the biggest weaknesses. In its global threat review, the report also noted that attackers are increasingly relying on AI to support phishing, social engineering and credential harvesting.
The cloud bill, however, is only one part of the risk for LLMJacking. Depending on the permissions attached to compromised credentials, an attacker can elevate privileges and move laterally to other systems, stealing source code and sensitive data, encrypting data and disrupting services. A cloud provider that detects suspicious activity may suspend the account and this then impacts customer service. There is also a regulatory concern because, under Kenya’s Data Protection Act, if personal data has been accessed by an unauthorised person and there is real risk of harm, the data controller must notify the Data Commissioner within 72 hours of becoming aware of the problem.
Another part of the problem is how difficult it is to detect LLMJacking. The attacker is using verified credentials and works to blend in with normal activity. If your business was running 10 virtual machines, for example, you’d notice an 11th but if you were running hundreds, you may not notice a few more instances. Google found that in several incidents the victim had not enabled LLM services and the attackers switched them on. One group tested stolen keys with requests limited to a single output token to stay below volume-based alerts because detection tends to rely on watching for sudden spikes in usage.
Knowing the risk is one side of the story, the other is protecting against it. You can build a defence that aligns with the threat by prioritising several key steps. First, check your repositories before you commit anything to a public repository and ensure you sanitise the code so there are no sensitive keys or tokens. Rotate your keys regularly so a potentially compromised key will have a very limited lifespan. Protect your logins with multi-factor authentication (MFA) and keep your development, staging and production environments separate with different keys so one compromised key doesn’t open every environment. When you segregate your environments, it makes it difficult for one malicious actor to transverse or laterally move from one to another.
Visibility is another part of the solution. Enable billing alerts and spending caps on your cloud accounts and review your audit logs for permission changes or unfamiliar logins. It’s also key to treat any unexpected activation of AI services as a security event. Revoke any credentials that may have been exposed as well, issuing new ones regularly, and always be aware of where your keys are, who can use them and what normal activity looks like in their cloud accounts.
The practical lesson for Kenyan companies is to manage cloud keys and API tokens with the same discipline as passwords and financial controls with named owners, limited lifespans and regular reviews so you can stay ahead of the LLMJacking risks and any potential vulnerabilities.
