Stolen AI API keys generated nearly $1M in charges for a single company (cloud.google.com)

Mandiant's recent report reveals a troubling trend in cybersecurity known as "LLMJacking," where threat actors hijack access to large language models (LLMs) to generate fraudulent charges exceeding $1 million for affected companies. This has been observed across various sectors globally, with attackers exploiting compromised cloud resources and LLM API keys. Unlike traditional cryptojacking that incurs gradual resource consumption, LLMJacking results in rapid spikes in charges, as adversaries utilize stolen credentials to run high-volume operations, often in just hours. The significance of LLMJacking is underscored by the sophisticated underground market that has emerged, characterized by distinct roles among attackers, including those who gain initial access and those who monetize the stolen resources. Procedures to compromise systems involve targeting exposed LLM endpoints or stealing cloud identity tokens, leading to extensive privilege escalation and subsequent use of hijacked API access for illicit model usage. Threat actors have created proxy infrastructures to manage and sell these resources at reduced costs, further complicating the security landscape for AI/ML providers and users alike. The report emphasizes the urgent need for organizations to bolster their defenses against these evolving tactics in the AI/ML domain.