Affected Systems
Organizations using cloud-hosted large language model (LLM) services from Google and other providers. Attacks target exposed API keys, misconfigured cloud resources, GitHub tokens, service account keys, and session tokens. LiteLLM users affected by CVE-2026-40217 exploitation enabling arbitrary code execution.
Exploitation Status
Active exploitation confirmed. Google Threat Intelligence Group and Mandiant observed multiple threat clusters including UNC6993. Attackers exploit CVE-2026-40217 in LiteLLM and operate LLM mixer services (e.g., Jet Night, newapi[.]pro) that resell stolen access. Request volumes exceed 1,000 requests/second in observed incidents.
Business Impact
Organizations face fraudulent cloud charges exceeding $100,000 per day, with cumulative losses reaching $1 million at some victims. Attackers scan for exposed LLM endpoints, steal API keys, and resell access through proxy services charging victims for token consumption. Threat actors establish persistence via backdoor accounts, SSH key injection, and unauthorized proxy infrastructure on victim cloud environments. Shadow AI services complicate detection.
Urgency
🟠Within 24 hours
Recommended Actions
- Implement centralized secrets management platform with automated rotation for API keys, service account keys, and LLM credentials
- Enforce phishing-resistant MFA on all cloud accounts and deny permissions not explicitly required, including access to Cloud LLM services
- Enable cloud audit logs and LLM inference logs; monitor for unauthorized service activation, suspicious credential checks, unusual usage spikes (>1000 req/sec), and logging disablement attempts
- Set API usage rate limits and spending caps on all LLM services to contain financial impact of compromise
- Patch LiteLLM instances to address CVE-2026-40217 and scan GitHub repositories, logs, databases, and secret managers for exposed credentials