Skip to main content
All incident reports
Incident report

PoeLLM malware infects 3,400 exposed AI servers to mine cryptocurrency

A cryptomining campaign is using malware called PoeLLM to compromise exposed AI and developer services, then turn them into scanners that hunt for more victims. Lumen's Black Lotus Labs says more than 3,400 servers in the US and Western Europe have been infected since April 2026.

· First published

The short version

  • The campaign. Lumen's Black Lotus Labs, which named the malware PoeLLM and the campaign Canto Incognito, says more than 3,400 servers have been compromised since at least April 2026, concentrated in the US and Western Europe.
  • What it does. Infected hosts run the XMRig and Iron cryptocurrency miners, connect to the Russian mining service Kryptex, and are repurposed to scan for further targets and spread the malware.
  • What is targeted. The campaign goes after internet-facing LiteLLM, Ollama, Gotenberg and Gitea deployments, with signs of Ivanti Sentry targeting as well. Lumen says exposed AI services appeal to attackers because they are often poorly configured and run on powerful GPU hardware.
  • Attribution. Lumen says it cannot attribute the activity confidently but assesses with moderate confidence that the operator is Italian-speaking, based on comments in the malware and Italy-based infrastructure. The malware uses a poem on GitHub to work out where to connect.

Timeline

DateEvent
13 AprilFirst commit is made to the GitHub repository that hosts the poem the malware reads.
AprilPoeLLM activity begins; researchers say the campaign has been active at least since this month.
Mid-JunePeak of operations: almost 2,200 affected servers, with nearly 800 active per day.
7 OctoberBlack Lotus Labs publishes its report; the victim count is revised from 2,100 to more than 3,400 in the live document.

How the infection works

PoeLLM is an ELF file named libgcrypt. To find its command-and-control server, it reads four words or phrases from a poem titled "On the Nature of Connection" held in a dash.css file in a GitHub repository that appears to fork Node.js, then converts them into an IPv4 address using a hard-coded dictionary. The operator changes the poem to move the address; Black Lotus Labs says it has been modified 11 times and suspects at least one more update. The repository's first commit was on 13 April 2026.

Once installed, the malware provides a remote shell, the XMRig and Iron cryptocurrency miners, HTTP and HTTPS scanning, and the ability to deploy exploits. Infected machines scan ports 3000 and 4000, associated with Gotenberg and LiteLLM, and attempt to exploit CVE-2026-42271, a flaw in LiteLLM's MCP server test endpoints that was first disclosed as requiring authentication. Horizon.ai researchers found it could be chained with a second issue, CVE-2026-48710, to reach unauthenticated remote code execution.

Who is affected

The victims are enterprise systems reachable from the internet: LiteLLM and Ollama AI tools, the Gotenberg PDF converter, the Gitea development toolkit, and Ivanti Sentry appliances where Lumen found signs of targeting. No victim has been named publicly, and neither article carries a statement from an affected organisation confirming an intrusion.

Lumen puts the number of compromised servers above 3,400. BleepingComputer reported an initial figure of 2,100, which the researchers changed to 3,400 in their live report. At the mid-June peak the campaign involved almost 2,200 affected servers, with nearly 800 active on a single day, and at least 11 command-and-control servers have been set up so far.

Attribution and open threads

Black Lotus Labs says it could not make a confident attribution, but assesses with moderate confidence that the operator is Italian-speaking. BleepingComputer cites Italian-language comments in the malware and an Italy-based server hosting an administrative interface; The Hacker News cites Italian-language artefacts and netflow indicators as the basis.

Several of the command-and-control servers had vulnerable router administration interfaces, which the researchers say suggests the attacker reused compromised routers rather than building fresh infrastructure. More recent traffic toward SSH and other login portals points to experimentation with distributed brute-force attacks, but Lumen says how mature that capability is remains uncertain.

What is not known yet

  • Who is behind the campaign; Lumen has only moderate confidence that the operator is Italian-speaking.
  • Whether any data was taken from the compromised AI and developer services, beyond their compute power being abused.
  • How many of the 3,400 servers are still infected and mining at the time of publication.
  • Whether the distributed brute-force attempts against SSH and login portals have become a working capability rather than an experiment.
  • Which organisations own the affected servers; no victims have been named.

What to do

If you run LiteLLM, Ollama, Gotenberg, Gitea or Ivanti Sentry

  1. Apply the latest security updates, including the LiteLLM fixes for CVE-2026-42271 and CVE-2026-48710.
  2. Restrict external access to trusted IP addresses and take any service that does not need to be public off the internet.
  3. Check from outside your network whether ports 3000 and 4000 are reachable.

If you suspect a server is compromised

  1. Look for a file named libgcrypt and for XMRig or Iron miner processes.
  2. Search network logs for connections to the indicators of compromise Black Lotus Labs published and to the Kryptex mining service.

For everyone else

  1. Inventory internet-facing AI and large language model services and confirm that each one needs to be publicly reachable.
  2. Monitor for outbound scanning from internal hosts, which indicates a machine has been turned into a scanner.
  3. Treat publicly exposed services on powerful hardware as a mining target, not only a data risk.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.