Cybercriminals Exploit Rising AI Costs with Growing LLM Jacking Attacks on Organizations

As organizational spending on AI models has surged recently, cybercriminals have increasingly exploited this trend through a cyberattack known as LLM Jacking. This attack involves stealing API keys that grant direct access to large language models (LLMs), enabling hackers to misuse AI resources and cause substantial financial damage. Experts estimate that cumulative losses from such attacks have reached around one billion dollars this year alone.

Dima Tatur, Vice President of Cybersecurity at Commit, described LLM Jacking as the hottest AI cybersecurity threat currently. Attackers can inflict hundreds of thousands of dollars in damages within a week to ten days before organizations even realize the breach. The attacks have grown in frequency alongside the rising costs and expanded use of AI in enterprises.

The vulnerability arises because AI agents and organizational interfaces typically connect directly to AI models via API keys, which act as authentication tokens. Many organizations inadvertently leave these keys exposed in application code or public repositories like GitHub. Attackers then steal these keys, gaining unrestricted access to the organization's AI "wallet." Rather than using the stolen keys themselves, criminals often resell access through dark web token marketplaces, offering cheaper AI usage to other buyers, including both illicit actors and unwitting legitimate companies seeking cost savings.

The financial impact on victim organizations can be severe, with daily losses averaging $50,000 to $100,000 and some accounts incurring nearly $400,000 in monthly charges. Beyond direct costs, organizations face risks such as service shutdowns when usage limits are exceeded and exposure of sensitive organizational data accessible through the compromised AI sessions.

To mitigate these risks, Tatur recommends using virtual API keys that provide better monitoring and real-time alerts. In response to an attack, organizations should immediately block compromised keys, securely replace them without embedding them in code, and investigate the breach path. However, cooperation from AI model providers like OpenAI and Anthropic remains limited, complicating incident management and recovery efforts.