PoeLLM malware infects more than 3,400 AI servers for mining

1. 18:24
   Help Net Security AI

   PoeLLM botnet hides C2 address via GitHub poem, infecting more than 3,400 servers

   Black Lotus Labs discloses a mining botnet named Canto Incognito, whose malware PoeLLM resolves C2 server addresses by reading a GitHub poem titled On the Nature of Connection; the poem has been modified 11 times since its first commit in April 2026, each time pointing to a new server. The campaign has been running since April 2026 and has infected more than 3,400 servers, most located in the United States and Western Europe, with victims mostly running vulnerable versions of open-source AI/LLM services such as LiteLLM and Ollama, as well as Gotenberg and Gitea. Black Lotus Labs says it has blocked traffic to PoeLLM C2 servers and continues to monitor.

2. 00:01
   网页The Register· Discussion

   Poetry is the new AI security threat as PoeLLM malware infects 3K+ servers(opens the original post in a new tab)

1. 23:33
   The Hacker News

   PoeLLM malware infects more than 3,400 servers, expanding a cryptocurrency mining botnet

   Lumen Black Lotus Labs disclosed a malicious campaign named Canto Incognito, whose delivered PoeLLM malware has infected more than 3,400 AI and LLM infrastructures exposed to the internet, used to deploy cryptocurrency miners such as XMRig and Iron and connect to the Russian mining service Kryptex. The attacks primarily target enterprise external deployments such as LiteLLM, Gotenberg, Gitea, and Ivanti Sentry, with infections concentrated in the United States and Western Europe. The malware hides the command-and-control address in the words of a poem in a GitHub repository; each time the C2 changes, several words in the poem are altered, and the address is derived from the corresponding keys. Infected hosts are also turned into scanners and exploit servers, sending HTTP POST requests to exposed ports to lure downloads of the malware, expanding the botnet.

2. 23:00
   CyberScoop

   PoeLLM malware hides C2 address via GitHub poem, has compromised more than 3,400 servers

   Lumen Technologies' Black Lotus Labs reports that a malware named PoeLLM has compromised more than 3,400 servers since April, primarily targeting open-source AI services, and uses a poem published on GitHub to hide the command-and-control (C2) server address. The poem appears harmless; the malware extracts four specific words from the poem based on fixed text anchors, then maps each word to a set of IP addresses via a built-in dictionary, and the four numbers combine to form the current C2 address. Threat actors can change the C2 location simply by rewriting the poem, without updating the malware itself; the relevant keywords have been changed at least a dozen times. Researchers first discovered the PoeLLM infrastructure in June this year while investigating a maximum-severity vulnerability in Ivanti Sentry, and subsequently uncovered a large-scale vulnerability scanning and cryptocurrency mining botnet involving compromised services and tools such as LiteLLM, Ollama, Gotenberg, and Gitea.

3. Time not given
   Lumen Black Lotus Labs

   Black Lotus Labs discloses the PoeLLM malware campaign; more than 3,400 servers compromised

   Black Lotus Labs has been tracking a malware campaign named PoeLLM since April 2026, which targets AI and open-source services exposed to the internet to deploy cryptocurrency miners and scan for new targets. The malware primarily affects vulnerable external deployments such as LiteLLM, Ollama, Gotenberg, and Gitea, and may also involve Ivanti Sentry. Its command-and-control mechanism is unusual: it extracts four keywords from a poem titled "On the Nature of Connection" in a GitHub repository, converts them into numbers via a hardcoded dictionary, and assembles the current C2 IPv4 address. The poem has been modified 11 times since its initial commit on April 13, each time pointing to a new C2. The campaign has affected more than 3,400 victim servers, with over 800 active servers per day at its peak; researchers say the peak in mid-June involved nearly 2,200 affected servers.

Current attention41·Peak in the comparable period56（Oct 8 02:00）·Change over the comparable past 24 hours 0%