PoeLLM campaign reportedly turns exposed AI servers into cryptomining and attack infrastructure

Retrieving command addresses from keywords in a poem

According to the report, PoeLLM is malware that retrieves command-and-control addresses from keywords embedded in a poem hosted in a GitHub file. Black Lotus Labs says the botnet has compromised more than 3,400 servers, with many affected systems running exposed tools including LiteLLM and Ollama. The researchers describe remote-shell, cryptocurrency-mining, scanning, and exploit-deployment capabilities.

The supplied report does not establish a directly comparable prior count of affected servers or a baseline for the campaign's current activity.

The reported campaign is described as using compromised servers not only for cryptomining but also for scanning and attempts to spread to further systems.

The reported operation combines cryptomining with reuse of compromised AI-service infrastructure for discovery and exploitation of additional targets. Its command-and-control lookup method, based on text in a poem, is also an unusual technique described by the researchers.