PoeLLM Malware Hijacks Exposed AI Servers for Cryptomining
What Happened — A new cryptomining botnet dubbed PoeLLM has compromised more than 3,400 AI‑focused servers worldwide. The malware leverages poorly‑configured, internet‑exposed AI services (e.g., LiteLLM, Ollama, Gotenberg) and uses a novel “poem‑based” C2 address resolution to control infected hosts. Infected machines run XMRig miners, act as scanners, and attempt to exploit CVE‑2026‑42271 (and chained CVE‑2026‑48710) to expand the foothold.
Why It Matters for Trust & Control Assurance
- Highlights the need for continuous configuration‑management monitoring to detect and remediate unintentionally exposed AI endpoints.
- Demonstrates how undocumented assets become evidence gaps in audit trails; continuous control‑assurance tooling can capture configuration drift and unauthorized exposure.
- Shows that a single misconfiguration can cascade into a supply‑chain‑like cryptomining operation, underscoring the importance of real‑time logging and anomaly detection.
Who Is Affected – Cloud‑native AI service providers, SaaS platforms offering LLM APIs, and any organization exposing GPU‑powered inference servers (tech‑SaaS, cloud‑infra, research labs).
Recommended Actions –
- Inventory all publicly reachable AI/LLM services and enforce strict network‑level access controls.
- Deploy continuous configuration‑validation tools that flag exposed ports (e.g., 3000, 4000) and missing authentication.
- Integrate real‑time telemetry from these assets into a control‑assurance dashboard to maintain a defensible audit trail.
Technical Notes – The botnet uses an ELF payload (libgcrypt) that parses a poem hosted on GitHub to generate C2 IP addresses. It spreads via port scans on typical AI service ports and exploits CVE‑2026‑42271 (LiteLLM MCP endpoint) and CVE‑2026‑48710 (remote code execution). Communication is routed through Russian crypto‑mining service Kryptex; compromised router admin interfaces are reused for C2 hosting. Source: BleepingComputer