John Hultquist, chief analyst of the Google Threat Intelligence Group, has raised concerns in an interview with the Financial Times about a growing threat known as LLM-jacking. This technique allows hackers to access high-end artificial intelligence (AI) models at a fraction of their usual cost. While the practice is not new, Hultquist noted a significant rise in its use in 2026, as illicit access to AI has become one of the most valuable commodities in the underground cybercrime market, sought after by both individual hackers and advanced persistent threat (APT) groups—often linked to state-sponsored activities. Hultquist described two main ways LLM-jacking is being carried out. In one case, hackers steal login credentials for popular AI platforms and sell access to these accounts on dark web marketplaces. Buyers can then use high-end subscriptions, such as Claude Max or ChatGPT Pro, which typically cost around $200 per month, for a much lower price. The original account holder remains unaware that someone else is using the subscription, continuing to pay the full cost. Hultquist mentioned that access to models from companies like Anthropic, Google, and OpenAI is sometimes sold at discounts as high as 97 percent. Some sellers even offer "guaranteed access," providing new credentials for free if the first one is blocked, to reassure buyers against detection by AI companies that monitor for abuse. In the second scenario, there is no resale of stolen access. Instead, more organized criminal groups—sometimes backed by states—hack into enterprise cloud servers that companies use to run their services. Once inside, they install their own AI models on these servers. The hacked company ends up paying for the computational power used, while the attackers gain significant capacity without any cost. This method is similar to cryptojacking, where hackers hijack third-party machines to mine cryptocurrencies. Google teams have observed this tactic used by criminal groups, as well as by a Chinese cyberespionage group previously linked to attacks on the United States. Hultquist did not provide further details about this group. Hultquist warned that this second method could become more common as large companies increasingly choose to host their own AI models rather than renting computing power from cloud providers. These expensive infrastructures will become prime targets, especially since companies are still figuring out how much computing power they need. A sudden increase in usage could be mistaken for normal activity following a recent deployment, making it harder to detect unauthorized access.