Observed incident

Mandiant documents enterprise LLM access theft and illicit resale

Reported

Published here

What happened

GTIG investigated stolen cloud/model access, persistence and inference resale. Some organizations accrued about $1 million in unauthorized usage charges.

Parties and reporting

AI tools and systems: Cloud LLM APIs, LiteLLM gateways, Manus-based backdoor, LLM resale proxies

Outcome

Identity compromise and billing abuse were observed; campaign-wide containment and final paid losses remain unknown.

Operating lesson

Restrict model permissions, rotate exposed identities and alert on new keys and abnormal inference spending.

Evidence and limits

Grouped 2026 investigations; victim counts and overlap with prior breaches are unknown. Not every intrusion used agents. GTIG labels user-ID rebinding as CVE-2026-40217, but the CNA identifies guardrail code execution. That exact mapping remains unconfirmed.

Sources