A financially motivated campaign dubbed Canto Incognito is targeting exposed AI and LLM infrastructure with a new malware family called PoeLLM, turning compromised servers into cryptomining nodes, scanners, and exploit launchpads. Lumen Black Lotus Labs reports more than 3,400 victim servers, mostly in the US and Western Europe, with peaks of approximately 800 active systems per day. PoeLLM (an ELF binary masquerading as libgcrypt) uses a creative command-and-control (C2) mechanism that reads four (4) keywords from a poem (“On the Nature of Connection”) in a dash.css file on a GitHub repo, maps them via a hard‑coded dictionary to an IPv4 address, and thus derives the C2. Each time the operator changes the poem, the botnet points to a new C2, with at least eleven (11) having been used so far. The malware deploys XMRig and Iron miners, provides remote shell access, scans HTTP/S, and exploits exposed services like LiteLLM, Ollama, Gotenberg, Gitea, and Ivanti Sentry, including chaining LiteLLM vulnerabilities CVE‑2026‑42271 and CVE‑2026‑48710 for unauthenticated remote code execution (RCE). Compromised hosts are reused to scan ports 3000/4000 and push the malware to new targets, and some C2 servers appear to run on previously hacked routers. Lumen assesses with moderate confidence that the actor is Italian, based on code comments and netflow to an Italy‑based admin server. CTIX Analysts recommend that organizations and defenders patch affected services, reduce internet exposure for AI/LLM endpoints, restrict access to trusted IPs, and hunt for connections to BLL’s published IoCs.