Stolen LLM API tokens driving illicit charges above $100,000 a day
Original title: Token thieves costing companies $100K+ per day in illicit LLM charges
Mandiant says some victims reached about $1 million before cutting access; one intruder made over 500,000 calls in under eight hours.
Google’s Mandiant says criminals are hijacking corporate LLM API access, a practice it calls LLMJacking, and generating token bills that can exceed $100,000 a day. UNC6993 used stolen keys, created more keys and enabled services in a victim cloud project. Other intruders subscribed accounts to pay-as-you-go plans and rotated short-lived service accounts. Google also describes Jet Night, a low-priced token-routing service it believes depends on stolen keys.
Unconfirmed: Google believes Jet Night is profitable because it uses stolen keys; that dependence is not independently confirmed.
- A CrowdStrike case saw about 200,000 API requests in two minutes.
- UNC6993 entered in mid-2026 through a misconfigured service-account key.
- A fake portal asked users to paste JSON access tokens, bypassing passwords and MFA.
- Jet Night advertises unlimited tokens for about $60 a month.
- Google says initial-access crews and downstream “LLM Mixers” now split the work.
Impact: Runaway LLM charges from stolen cloud API access
- 1.Rotate LLM and cloud API keys with centralized secrets management
- 2.Limit users and service accounts to least-privilege permissions
- 3.Scan codebases and remove embedded API keys and secrets
- ATT&CK
- T1528 Steal Application Access TokenT1078.004 Valid Accounts: Cloud AccountsT1550.001 Application Access Token
- Threat actors
- UNC6993
- Organizations
- GoogleMandiantCrowdStrike
From the original article
3 min read at theregister.comVictim organizations are losing millions of dollars in token charges as criminals hijack their corporate LLM access to steal expensive computing resources and AI services.
By Jessica Lyons · theregister.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.