Threat research disclosures published today reveal a novel malware family abusing artificial intelligence server infrastructure. Threat actors chain Model Context Protocol test endpoint flaws with authentication bypasses to execute arbitrary system code and deploy persistent mining payloads.
The malware campaign, dubbed PoeLLM, targets exposed LiteLLM and Gotenberg AI servers. Adversaries chain LiteLLM MCP test endpoint vulnerability CVE-2026-42271 with unauthenticated access flaw CVE-2026-48710 to achieve remote code execution. PoeLLM utilizes a unique command and control C2 mechanism that translates poetic text blocks into IP addresses, deploying XMRig miners and reverse shell scripts across compromised AI build nodes. Threat telemetry confirms compromised servers actively scan internal subnets on ports 3000 and 4000.
Subverting artificial intelligence model serving nodes introduces compute hijacking and software supply chain risks. Because AI servers maintain high-performance GPU resources and API integration tokens, an unauthenticated host compromise permits threat actors to exhaust compute resources and harvest model integration keys.
– Apply official security patches published by LiteLLM maintainers across all AI server deployments immediately.
– Restrict public network visibility of LiteLLM test endpoints and Model Context Protocol ports using zero trust proxies.
– Inspect AI host process trees for unauthorized XMRig execution routines or anomalous port scanning activity.
– Rotate all API integration tokens and model repository credentials processed on exposed AI application servers.
AI infrastructure defense relies on strict endpoint authentication and continuous process monitoring to ensure machine learning build nodes remain protected against unauthenticated exploit chains. #CodeDefence #LiteLLM #AISecurity #PoeLLM #RCE #CryptoMining #AppSec #DevSecOps #PatchManagement
/
