Artificial Intelligence · Cybersecurity
PoeLLM malware hit exposed AI servers in cryptomining campaign
Lumen's Black Lotus Labs said PoeLLM infected more than 3,400 servers, used a poem on GitHub to find command-and-control addresses, and targeted exposed AI tools and other services.
- Lumen's Black Lotus Labs said PoeLLM had compromised more than 3,400 servers and reached about 800 active infections on one day.
- The campaign targeted exposed services in the United States and Western Europe, including LiteLLM, Ollama, Gotenberg, Gitea and Ivanti Sentry.
- BLL said the malware read words from a poem in a GitHub-hosted file and mapped them to an IPv4 address for its control server.
- The researchers said they had seen links to Kryptex and could not firmly identify the operator, though they assessed the operator as likely Italian; Horizon.ai said CVE-2026-42271 could be chained with CVE-2026-48710.
Black Lotus Labs said PoeLLM had compromised more than 3,400 servers, with as many as 800 infected machines active on one day. It said the malware had been active since at least April and that the number of control servers had reached 11.
The campaign went after exposed AI-related systems in the United States and Western Europe. The services named in the report included LiteLLM, Ollama, Gotenberg, Gitea and signs of Ivanti Sentry.
BLL said PoeLLM was an ELF file called libgcrypt that recovered its control address by pulling words from a poem in a dash.css file on GitHub. Those words were translated through a fixed lookup table into an IPv4 address, and the poem was altered whenever the operator wanted a new control server.
The malware also carried remote-shell functions, XMRig and Iron miners, scanning over HTTP and HTTPS, and tools for exploit delivery. BLL said victims connected to Kryptex, that some control servers appeared to reuse compromised routers, and that the operator was probably Italian, although the attribution was not firm. The researchers advised patching systems, limiting public exposure and checking logs for their indicators of compromise.
Named in this story
Companies
- Kryptex
- was the crypto-mining service seen in victim communications
Organisations
- Lumen's Black Lotus Labs
- tracked the botnet and reported the findings
- Horizon.ai
- said one LiteLLM issue could be chained for remote code execution
Products and systems
- PoeLLM
- the malware family used in the campaign
- LiteLLM
- was among the exposed AI services targeted
- Ollama
- was among the exposed AI services targeted
- Gotenberg
- was among the services scanned and targeted
- Gitea
- was among the services scanned and targeted
- Ivanti Sentry
- showed signs of being targeted
- XMRig
- was included as a cryptocurrency miner
- Iron
- was included as a cryptocurrency miner
Places
- United States
- was part of the target region
- Western Europe
- was part of the target region
- Italy
- was tied to the suspected operator and one C2 host
How the source tells it
The piece was mostly reportorial, with urgency from the scale of the botnet and novelty from the poem-based command lookup.
- urgency the large compromise count and the reported daily peak
- novelty an unusual method for deriving command-and-control addresses from a poem