In its October 2026 analysis, Lumen Black Lotus Labs describes PoeLLM, a campaign it has tracked since April that compromises internet-accessible AI services and other applications. Operators install cryptocurrency miners; the malware derives its command-and-control server IP from words in a poem hosted on GitHub.
Exposed services as an entry point
Lumen says operators search for AI generation or inference services that are accessible without adequate protection. The campaign also targets tools such as Gotenberg and Gitea. Exposure alone does not prove compromise: logs and process activity need to be reviewed.
A payload focused on mining
Lumen’s analysis describes cryptocurrency miners and persistence mechanisms. Command traffic can be encoded or obscured, so the researcher’s technical indicators are more useful than looking for a single filename. Secondary reports repeat this analysis and do not independently establish a victim count.
Reduce exposure of AI interfaces
Inventory APIs and admin consoles, remove unused interfaces from the public internet, and enforce strong authentication and network rules. Patch exposed components and restrict service-account permissions. A gateway does not replace network isolation.
Look for signs of compromise
Check for unexpected mining processes, scheduled jobs, unusual outbound connections, and container changes. Preserve logs before remediation and compare indicators with Lumen’s reporting. If unauthorized access is confirmed, rotate secrets reachable from the host after containment.
How Soclyde fits
For teams operating AI services, Soclyde can help protect API credentials and share them with authorized people. It does not monitor hosts or mining payloads. See our guide to unique passwords and contact our team.
What to remember
Inventory internet-facing interfaces and review processes and outbound connections; Lumen’s published victim counts are inconsistent across sections of its analysis. For account hygiene, see our unique-password guide.
