Artificial intelligence is becoming an increasingly useful tool for small and medium-sized businesses. From automating everyday tasks and improving customer service to analysing data and supporting software development, large language models (LLMs) are helping businesses work more efficiently. However, as organisations become more reliant on AI services, a new cybersecurity risk is emerging: LLM jacking.
Unlike attacks that target AI systems themselves, LLM jacking involves cybercriminals gaining unauthorised access to an organisation's AI accounts, API keys or cloud infrastructure to use its AI resources for their own purposes. The consequences can include unexpected costs, compromised business information and disrupted operations.
For SMBs that may not have dedicated IT or security teams, understanding this threat is an important part of adopting AI responsibly. Here's what LLM jacking is, how it works and what your business can do to reduce the risk.
What is LLM jacking?
LLM jacking is a type of cyberattack in which criminals steal or exploit access credentials to use someone else's large language model services without permission.
LLMs are the technology behind many popular generative AI tools. Businesses often access these models through cloud platforms or application programming interfaces (APIs), which allow different applications and systems to communicate with an AI service.
These services are frequently billed according to usage, such as the number of tokens processed or the computing resources consumed. This creates an opportunity for cybercriminals: if they can gain access to a business's AI account or API credentials, they can use its resources while the legitimate account holder receives the bill.
In some cases, criminals use stolen credentials to run their own AI workloads. They may also resell unauthorised access to other users, effectively turning a compromised business account into a source of income.
For an SMB, this means that an AI service intended to improve productivity could become an unexpected financial liability and a potential entry point for further cyberattacks.
How does LLM jacking work?
LLM jacking typically starts with compromised credentials or an insecurely configured AI service. Cybercriminals use several methods to gain access, many of which exploit weaknesses in everyday business cybersecurity practices.
Common attack methods include:
- Stolen API keys: API keys are used to authenticate applications and connect them to AI services. Criminals may find exposed keys in public code repositories, unsecured files, application code or shared documents.
- Phishing attacks: Employees may be tricked into entering their login details on fraudulent websites or sharing authentication information through convincing emails and messages.
- Compromised accounts: If an employee's cloud account or AI platform credentials are stolen, attackers may be able to access connected services and resources.
- Insecure cloud configurations: Poorly secured cloud environments, excessive user permissions and publicly exposed AI endpoints can give attackers opportunities to access AI infrastructure.
- Compromised devices: Malware or other malicious software installed on an employee's computer can steal saved credentials, session tokens and other sensitive information.
Once inside, attackers can use the compromised credentials to access AI models, generate large volumes of requests and consume computing resources. They may also create additional access keys, change account settings or attempt to expand their access to other cloud services.
Because the activity can initially resemble legitimate AI usage, unauthorised activity may go unnoticed until unusual usage patterns, performance issues or unexpected billing alerts appear.
How does LLM jacking affect small businesses?
While large enterprises may have the resources to absorb unexpected cloud expenses or investigate complex security incidents, the same cannot always be said for smaller businesses.
LLM jacking risks for SMBs extend beyond the immediate cost of unauthorised AI usage. A successful attack can affect several areas of the business.
Unexpected costs and financial losses
One of the most immediate consequences is an increase in AI usage charges. Since many AI services operate on a pay-as-you-go model, attackers can generate substantial costs by making repeated requests or accessing more expensive models.
In some cases, cybercriminals use compromised accounts to run automated workloads continuously or resell access to stolen AI resources. This can result in unexpectedly high bills before the business realises its account has been compromised.
Exposure of sensitive business information
Access to AI services may provide attackers with opportunities to access conversation histories, uploaded documents, API integrations or connected cloud resources, depending on the permissions and configuration involved.
For SMBs, this could put customer information, financial records, internal documents, intellectual property and confidential business communications at risk.
It is also worth remembering that employees may enter sensitive information into AI tools without fully understanding how that information is stored, processed or shared.
Operational disruption
When cybercriminals consume an organisation's AI resources, legitimate users may experience slower services, usage restrictions or unexpected interruptions. Access may also be suspended if unusual activity triggers security controls or billing limits.
For businesses that rely on AI-powered tools for customer support, administration or other essential operations, even temporary disruption can affect productivity and customer service.
Further cybersecurity risks
A compromised AI account can sometimes be a stepping stone to wider attacks. If the stolen credentials provide access to connected cloud services, applications or business systems, attackers may attempt to move beyond the original AI environment.
This highlights why AI security should form part of a broader cybersecurity strategy rather than being treated as a separate technology concern.
How to protect your business from LLM jacking
As AI adoption grows, businesses need to think carefully about how their AI tools, accounts and connected infrastructure are secured. Fortunately, several practical measures can help reduce the risk of unauthorised access and misuse.
1. Secure your API keys and credentials
API keys provide access to AI services and should be treated with the same care as passwords or other sensitive business credentials.
Never include API keys in public code repositories, unsecured documents or applications where they can be easily extracted. Instead, store them in a dedicated secrets management solution or secure environment.
Where possible, use short-lived credentials, rotate API keys regularly and revoke any keys that are no longer required. If a key is accidentally exposed, replace it immediately rather than assuming that nobody has accessed it.
2. Apply access controls and limit permissions
Not every employee or application needs access to every AI service or business resource. Apply the principle of least privilege by giving users and applications only the permissions they need to perform their specific tasks.
Use separate API keys for different applications, projects and environments, such as development and production. This helps limit the potential impact if one set of credentials is compromised.
Multi-factor authentication (MFA) should also be enabled on AI platforms, cloud accounts and developer tools wherever available. This adds another layer of protection if login credentials are stolen.
3. Monitor AI usage and set spending limits
Regularly review AI usage logs, account activity and billing information to identify unusual behaviour.
Look out for unexpected increases in API requests, unfamiliar access locations, unusual activity outside normal working hours or the creation of new API keys and accounts.
Set usage quotas, rate limits and spending alerts wherever your AI provider supports them. These controls can help prevent excessive consumption and alert you to potential misuse before costs escalate.
Monitoring is particularly important for businesses using multiple AI tools or cloud-based services, as unauthorised activity may otherwise be difficult to identify.
4. Secure employee devices and business endpoints
An employee's computer or mobile device can provide an entry point for cybercriminals looking to steal credentials or access connected business systems.
Keep operating systems, applications and security software updated, and use trusted endpoint protection to detect and block malware, suspicious activity and other cyber threats.
It is also important to secure devices used to access AI platforms, particularly those with saved credentials, access tokens or development tools.
Endpoint security cannot replace API key management or cloud access controls, but it plays an important role in reducing the risk of compromised devices becoming the starting point for an attack.
5. Establish clear guidelines for using AI tools
Employees should understand which AI tools are approved for business use and what information they are permitted to share.
Create simple guidelines covering the use of public and private AI services, the handling of confidential information and the approval of new AI applications.
Avoid entering customer details, financial information, passwords, intellectual property or other sensitive business data into AI tools unless the service has been assessed and approved for that purpose.
Providing regular cybersecurity awareness training can also help employees recognise phishing attempts, suspicious login requests and other tactics used to compromise AI accounts.
6. Review your cloud security and connected services
AI tools rarely operate in isolation. They may connect to cloud storage, customer relationship management (CRM) platforms, internal databases and other business applications.
Review these integrations regularly to understand what information AI tools can access and what actions they are authorised to perform.
Disable unused integrations, secure exposed endpoints and check that cloud accounts have appropriate access controls in place. Businesses using AI agents should also assess what files, systems and external services those agents can access or interact with.
These measures help reduce the potential impact of a compromised account and strengthen overall cloud security.
Strengthen your AI security with multilayered protection
As businesses adopt more AI-powered tools, their cybersecurity strategies need to evolve alongside the technology. Protecting against LLM jacking requires more than securing an AI account: it also means protecting the devices, applications and credentials that provide access to business systems.
ESET PROTECT Advanced offers SMBs a way to strengthen their endpoint security through multilayered protection against malware, ransomware, phishing and other evolving cyber threats.
Its AI Security capabilities include AI Behavioural Monitoring, which monitors how AI tools and agents behave on endpoints to help identify suspicious or unexpected activity, and AI Agent Security, which provides a dedicated layer of protection for autonomous AI agents and their interactions with external resources.
These capabilities can help businesses address emerging risks associated with AI adoption and compromised devices. However, endpoint protection is only one part of an effective defence against LLM jacking. Businesses must also implement secure API key management, appropriate access controls, usage monitoring and cloud security measures to prevent unauthorised AI access.
By combining endpoint protection with good credential management and responsible AI usage, SMBs can reduce their exposure to emerging AI threats while continuing to benefit from the technology.
Make AI security part of your cybersecurity strategy
AI offers significant opportunities for small businesses, but adopting new technology also means understanding the risks that come with it. LLM jacking is a reminder that cybercriminals don't necessarily need to break into an AI system itself. Sometimes, gaining access to the credentials that unlock it is enough.
With the right precautions, businesses can reduce the risk of unauthorised AI usage, protect sensitive information and avoid unexpected costs. Most importantly, they can continue to use AI with greater confidence, knowing that security is part of the process rather than an afterthought.