Hackers Are Hijacking Your Company's AI Models

Attackers are hijacking enterprise AI models and cloud accounts, a new threat Google's security team calls 'LLMJacking,' resulting in massive unexpected bills.

Google’s cybersecurity division, Mandiant, has identified a new and costly threat targeting companies that use artificial intelligence. Dubbed “LLMJacking,” this attack involves threat actors hijacking a company's access to large language models (LLMs) and the underlying cloud infrastructure. In a recent report, Mandiant detailed multiple incidents where attackers gained unauthorized access to enterprise AI resources across various cloud and LLM providers. The primary goal of these attacks is not to steal data but to exploit expensive computational resources for the attackers' own purposes. The financial consequences are severe, with some victim organizations facing unexpected bills ranging from tens of thousands to over a million dollars in fraudulent usage charges.

The mechanics of LLMJacking are straightforward and exploit common security oversights. Attackers typically gain access by finding exposed credentials, such as API keys or cloud service account tokens. These keys are often accidentally leaked in public code repositories on platforms like GitHub, left in misconfigured cloud storage buckets, or stolen from developer workstations. Once an attacker possesses these credentials, they have the keys to the kingdom. They can use the victim’s account to make expensive API calls to proprietary models like those from OpenAI or Anthropic, or they can spin up powerful and costly GPU instances in the victim's cloud account. These hijacked resources are then used for various unauthorized activities, including running their own AI-powered services, training models, or even mining cryptocurrency, all while the legitimate account holder foots the bill.

This emerging threat fits into a broader pattern of resource abuse attacks, representing a modern evolution of cryptojacking. For years, attackers have hijacked server CPU cycles to mine cryptocurrencies. LLMJacking applies the same principle to a far more valuable resource: AI compute. The rapid, widespread adoption of LLMs has created a new and lucrative attack surface. As companies race to integrate AI capabilities into their products, the security practices for managing the associated credentials have often failed to keep pace. This gap between rapid innovation and lagging security creates a perfect opportunity for attackers. The trend signifies a tactical shift for some cybercriminals, moving from data exfiltration to the direct theft and monetization of cloud services, where the payoff is immediate and tangible.

For any organization building with or using LLMs, this new threat demands an immediate focus on credential security and financial governance. LLM API keys and cloud credentials must be treated as highly sensitive secrets, equivalent to database passwords or private encryption keys. Teams should adopt a zero-trust approach, implementing strict access controls and ensuring that keys are never hardcoded in applications or committed to version control. Best practices include using dedicated secrets management tools, such as HashiCorp Vault or the native secret managers in AWS, GCP, and Azure. Furthermore, organizations must establish robust monitoring and alerting for both cloud spending and API usage. Setting up aggressive billing alerts and hard usage quotas can act as a critical safety net, preventing a minor credential leak from escalating into a million-dollar financial disaster. As AI becomes more deeply embedded in the enterprise, we can expect LLMJacking tactics to become more sophisticated, making proactive security and vigilant monitoring non-negotiable.

LLMJacking represents a new, financially devastating attack vector targeting the core of modern AI infrastructure. For developers and security teams, it highlights how unsecured API keys and cloud credentials for LLM services are now high-value targets for resource theft, not just data exfiltration.

The direct financial impact of LLMJacking can be severe, with Mandiant reporting incidents costing from $80,000 to over $1 million in unauthorized cloud and API usage. This threat introduces a significant new operational and financial risk for any company investing in or deploying LLM technology.