AI Account Lured by LLMJacking Surge: Watch for USD 100,000-a-Day Bill Shock
TECHWORLD ·
✦ AI Summary
LLMJacking is an attack in which stolen API keys are used to access AI models without authorization and resell that access at low prices to third parties.
GTIG and Mandiant said recent attacks occurred in North America, South America, and Asia-Pacific, with financial, retail, technology, and transportation companies among the targets.
The report proposed threat hunting, identifying shadow LLM use, segmenting API key privileges, and automated secret scanning as countermeasures.
LLMJacking attacks targeting companies' access to AI models are spreading. These attacks involve using stolen API keys to access AI models without authorization and then reselling that access at low prices to third parties based on the stolen API keys. As a result, companies face massive usage charges, increasing the need to strengthen AI resource management and security.
Google Threat Intelligence Group (GTIG) and Mandiant released a threat report on "LLM jacking" on the 4th. According to the report, recent attacks have occurred in North America, South America, and Asia-Pacific, and financial, retail, technology, and transportation companies have been identified as the targeted industries.
The report said LLMJacking is unlike CryptoJacking in that it does not gradually consume resources. Instead, LLMJacking is characterized by large-scale API calls, which lead to massive costs in a short period of time.
Abuse involving large language models has risen sharply in a short time. There have been cases in which more than USD 100,000 in abusive usage charges were generated in a single day, and in some companies, cumulative losses reached USD 1 million before unauthorized access was blocked.
Credential theft was cited as the initial access path for the attackers. They infiltrated targeted environments by using session hijacking, and they also collected valid credentials exposed in public GitHub repositories.
John Hultquist, senior analyst at GTIG, said attackers are aggressively using LLM jacking and seeking scale advantages over defenders. He also said attackers are using agentic tools to steal tokens without victims realizing it and expanding the scale of attacks on that basis.
Black market monetization and "LLM Mixers" were cited as the attackers' main revenue models. The attackers bundled stolen API keys into illicit proxy gateways, and they generated profits by reselling the unauthorized model access they secured at steep discounts in cybercrime markets.
One illicit token routing service offered buyers unlimited access to frontier models for about USD 60 per month, and in one case the resulting model usage costs were billed to the victim company at about USD 1 million. The case showed how the actual costs were passed on to the victim company after cheap unlimited access was offered.
Against this backdrop, the company said LLMJacking attacks are expanding and noted that the importance of corporate AI resource management is growing. As countermeasures, it proposed threat hunting for strategic threat detection, identifying shadow LLM use within organizations, segmenting API key privileges, and automated secret scanning.
John Hultquist, senior analyst at GTIG, said companies need to strictly control access to LLM resources to prevent cost shocks and protect the broader community from attackers' low-cost, autonomous operations. He said AI models and computing resources can be abused without authorization by third parties if they are not adequately monitored and protected, and recommended closely reviewing resource usage, enabling MFA, and carefully managing account information.
Source: TECHWORLD · Kim Hye-jin
Original: https://www.epnc.co.kr/news/articleView.html?idxno=409293
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.