PoeLLM

A profit-driven botnet found by black-lotus-labs, part of a wider operation the team calls Canto Incognito. Per gadgetreview-poellm-botnet: 3,400+ compromised servers, about 800 active on the peak day, running since at least April 2026 and still growing in October.

How it works

• Entry. Exposed AI and developer services: LiteLLM, Ollama, Gotenberg, Gitea, and one Ivanti Sentry appliance spotted when an infected unit scanned outward.
• Payload. XMRig and Iron cryptocurrency miners pointed at Kryptex mining infrastructure. GPU hosts are the prize, since inference hardware mines well.
• Spread. Infected hosts run scanners and exploits to recruit more victims.
• Control. The bot finds its command server through poem-c2-resolution, so the operator can move the server without touching the binary.

Attribution

Moderate confidence, a suspected Italian-speaking operator, from Italian-language artifacts and network-flow analysis. A GitHub account ejejejdfbbebe is named. The identity is unconfirmed.

Why the spoke keeps it

It is the spoke's first commodity-crime threat, and the first where the victims are AI serving infrastructure rather than AI as the attacker (compare emergent-agent-coordination). The lesson is ordinary: nothing here needed a new defence. Auth on exposed services, patching and egress monitoring cover it (system-hardening, network-traffic-analysis). Single secondary source, so no base rate for how common this pattern is.

Related

gadgetreview-poellm-botnet · poem-c2-resolution · black-lotus-labs · defensive-security