Researchers have described PoeLLM, a cryptomining-focused malware campaign that uses a GitHub-hosted poem to recover command-and-control information. The campaign reportedly affected more than 3,400 servers, with AI-serving and tooling environments among the prominent targets.
Using public content as a lookup mechanism can make simple blocklists less effective: malware can retrieve an innocuous-looking resource and derive current infrastructure from it. Reported targets include AI tools such as LiteLLM and Ollama.
Defensive actions
- Inventory externally reachable AI tools, inference endpoints and their supporting hosts.
- Review anomalous outbound requests, including unexpected GitHub retrievals and mining-pool traffic.
- Hunt for unauthorized containers, scheduled tasks and sustained high CPU use.
- Include AI tooling in normal patching, asset management and credential-rotation programs.
Public hosting is not inherently malicious; the detection question is whether a request occurs in an anomalous execution chain followed by suspicious compute or network behavior.
Source: Help Net Security, citing Lumen research (October 8, 2026).