In Other News: AI Used in Korean Bank Breaches, Poem-Guided Botnet, Empire Admin Gets 40 Years

What it is

This digest covers eight items, but the one with real technical substance the existing summary compresses too far is PoeLLM, a cryptomining botnet malware (active since at least April 2026, detailed by Black Lotus Labs) that targets exposed AI and open-source services — specifically LiteLLM, Ollama, Gotenberg, and Gitea.

How it works

Infected machines extract four keywords from a poem hosted on GitHub and convert those keywords into an IP address, which becomes the malware's command-and-control (C&C) server for that period. This gives the operator — assessed to be Italian-speaking — a simple way to rotate infrastructure: edit the poem's keywords, and every infected host resolves to a new C&C IP without any code update. The poem has already been updated 11 times.

Preconditions / Affected

Targets are exposed instances of LiteLLM, Ollama, Gotenberg, and Gitea — i.e., AI-serving and dev-tooling services reachable without adequate access control. The malware's purpose on compromised hosts is cryptocurrency mining and botnet expansion.

flowchart LR
    A[Exposed LiteLLM/Ollama/Gotenberg/Gitea] --> B[PoeLLM infects host]
    B --> C[Fetch poem from GitHub]
    C --> D[Extract 4 keywords]
    D --> E[Derive C&C IP]
    E --> F[Connect for mining/botnet tasking]
    G[Operator edits poem] -.rotates.-> D

What to do

The page gives no patch or IOC list for PoeLLM — no CVE, no specific version boundary, no indicator IPs or GitHub repo name. The actionable takeaway is architectural: these four service types should not be exposed to the internet without authentication, since that exposure alone is the entry point being exploited, not a software flaw being patched.

Caveats

Black Lotus Labs' writeup (as relayed here) does not name the GitHub repo hosting the poem, nor does it give the current or historical C&C IPs — so defenders can't build a detection rule from this digest alone; it establishes the technique, not indicators to act on directly.