LLMJacking Turns Stolen Cloud Credentials Into Six-Figure AI Bills

Attackers are stealing cloud credentials, LLM API keys, and access tokens to hijack corporate AI resources and resell access through discounted services, proxy gateways, and underground storefronts. Google’s Mandiant documented fraudulent charges exceeding $100,000 per day and victims accumulating nearly $1 million in costs before access was revoked. UNC6993 generated more than $100,000 in victim charges, while a separate operation supporting the Jet Night token-routing service imposed nearly $1 million in token costs on one organization. UNC6993 also used a deceptive application portal to obtain active access tokens, bypassing password and multifactor authentication controls.

Investigations found attackers expanding access through credential discovery, cloud privilege escalation, unauthorized accounts, SSH-key persistence, and exploitation of a LiteLLM gateway vulnerability tracked as CVE-2026-40217. Operators enabled premium models, rotated short-lived service accounts to evade rate limits, and used agentic workloads and GPU provisioning to scale abuse; Google assessed that stolen API quotas were being resold through services it calls LLM Mixers. Google recommends stronger identity and access management, centralized secrets storage and rotation, least-privilege permissions, and automated secrets scanning to reduce exposure.