Severity: High | Priority: 1
Google Threat Intelligence Group (GTIG) and Mandiant report a fast-growing financially motivated threat they call LLMJacking. Attackers take over an organisation’s cloud AI resources and model API keys, and that organisation pays for the attackers’ usage. They get in mainly with stolen cloud credentials, service-account keys, developer access tokens and API keys exposed in code repositories. They also use self-hosted model servers, such as Ollama, that are left reachable from the internet without authentication. Unlike cryptojacking, the costs build up within hours. GTIG documented fraudulent charges of more than USD 100,000 in a single day, and close to USD 1 million in total at some victims.
GTIG has found a resale market behind these attacks. Services it calls “LLM Mixers” pool stolen keys, meter usage and sell cheap access to underground customers. One service, Jet Night, made more than 500,000 API calls in under eight hours on stolen keys and passed nearly USD 1 million in costs to a single victim. Another actor ran relay infrastructure on lookalike domains to channel a victim’s model access into the newapi.pro resale ecosystem. Because newapi.pro is also associated with a legitimate open-source project, treat traffic to it as something to investigate, not as proof of compromise. In some cases attackers also targeted LiteLLM gateways, which centralise access to many model providers. They escalated ordinary virtual keys to administrator level, putting every credential behind the gateway at risk. CVE-2026-40217, a separate LiteLLM sandbox-escape flaw, is now in CISA’s Known Exploited Vulnerabilities catalog.
To protect your network, set hard spending limits and usage alerts on every cloud AI service, and get alerted whenever a new model or AI service is enabled. Give each API key the narrowest possible scope. Keep keys out of source code, and require phishing-resistant MFA for cloud administrators. Never expose Ollama or other model servers directly to the internet. Keep LiteLLM gateways patched, and block external access to their administrative endpoints.
If you see unexpected usage spikes, unfamiliar keys or administrator accounts, or traffic to known LLM Mixer services, revoke and rotate the affected keys right away and review cloud audit logs for further abuse.