Google: LLMjacking attacks rack up $1 million in cloud AI charges
AI Summary
Google's Threat Intelligence Group (GTIG) and Mandiant have identified LLMjacking attacks, where threat actors hijack access to enterprise large language model (LLM) infrastructure to generate fraudulent cloud AI usage charges, with some victims accruing over $1 million in costs. Attackers either exploit exposed LLM endpoints or steal valid API keys, then resell access through 'LLM mixers' like Jet Night and newapi[.]pro, which proxy user requests through compromised accounts. A specific threat cluster, UNC6993, has been observed aggregating stolen credentials and leveraging them at scale, including via exploitation of CVE-2026-40217 in LiteLLM to execute arbitrary code. Organizations are advised to strengthen IAM policies, secure secrets, enforce MFA, monitor logs, and set spending limits to mitigate such attacks.
AI-extracted · verify before operational use