For six months in 2026, thousands of hijacked servers kept checking a poem on GitHub to learn where their orders came from. The malware that read it, which Lumen's Black Lotus Labs calls PoeLLM, broke into exposed LiteLLM gateways, Ollama model servers, Gotenberg document converters and Gitea code hosts, installed cryptocurrency miners, and turned each victim into a scanner hunting for the next one. In a report published on October 8, 2026 the researchers put the total at more than 3,400 infected servers since April.
Self hosted AI tooling is being deployed at speed, often straight onto the public internet with no authentication, on machines with powerful GPUs and valuable API keys. The autumn of 2026 also brought two serious flaws in the same stack, an account takeover in LiteLLM and an unauthenticated remote code execution bug in LMCache, neither of which had a fix when it was disclosed. PoeLLM is a clear illustration of what that exposure costs.
Four words in a verse
PoeLLM carries no fixed command server address. It fetches a two stanza poem, On the Nature of Connection, and pulls four words from fixed positions in the text. A dictionary compiled into the binary converts each word to a number, and the four numbers form an IPv4 address; in one sample from June, driver, diode, decryption and string resolved to 92, 119, 165 and 74. To move the botnet the operator commits new words; the poem was edited 11 times after its first commit on April 13, 2026. Black Lotus Labs counted 12 command servers over the campaign's life, three still active when it published, several of them compromised routers rather than rented hosts. Italian comments in the code point to an Italian speaking operator.
Who got hit, and how
Scanning began in May 2026 and concentrated on ports 3000 and 4000, the defaults for Gotenberg and LiteLLM. A vulnerable host received a crafted POST request telling it to fetch a file from port 81 on the command server; the LiteLLM endpoint involved is the one referenced in the command injection flaw tracked as CVE-2026-42271, which Black Lotus Labs considers the likely way in. The payload, an ELF binary, bundles a remote shell and the XMRig and Iron miners, which it points at the Kryptex pool. At the peak in June almost 2,200 servers were affected, with nearly 800 active on a given day. One detail stands out: Gotenberg's own installation guide warns against putting the service on the open internet, and hundreds of servers were infected through it anyway.
The wider exposure problem
PoeLLM needed no novel tradecraft because the targets were already open. In September 2025 researchers at Cisco Talos used Shodan to look for Ollama servers reachable from the internet and found 1,139 reachable from the internet with no authentication at all. Only 214, about 18.8 percent, were serving live models; the rest sat idle but remained open to unauthorized model uploads and configuration changes. The United States hosted just over a third of them, with China and Germany next, and 88.89 percent exposed the standard OpenAI style API routes.
The software itself had a rough autumn. On September 29, 2026 OX Security disclosed CVE-2026-93355 in LiteLLM, an open source gateway to commercial model APIs, with a severity score of 8.8. When a signed login token does not match a known user, the gateway falls back to the unverified email address inside the token, so anyone with an account at a trusted identity provider can present a token carrying an administrator's email, be treated as that administrator, and permanently rebind the victim's account to their own identity. OX said it reported the bug on May 18, 2026, heard nothing back from the maintainers, and found it still present in LiteLLM 1.100.1 at disclosure.
A week later, on October 7, JFrog published CVE-2026-105192 in LMCache, a caching layer for the vLLM inference engine, with a score of 9.8. In multiprocess mode the software opens a ZeroMQ socket, by default on port 5555, with no authentication, and hands incoming messages to Python's pickle deserializer before any handler runs, so one message executes code as the LMCache user, which in the official container images is root. The socket stays on localhost unless an operator binds it to a routable address. JFrog said the vulnerable code shipped in version 0.3.9, was still in 0.5.5 and the 0.5.6 release candidates, and that no fixed release existed as of October 7, 2026.
What operators should do
- Never expose an inference API, a model gateway or a document converter to the internet without authentication in front of it; the Talos team's first recommendation is API key or OAuth based access control, or a VPN that keeps the service off the public address space.
- Segment and filter: put GPU hosts on their own network, keep internal transports such as LMCache's ZeroMQ port on the cluster network only, and restrict outbound traffic, since a server that can reach only its model registry cannot join a mining pool or fetch a payload from port 81. Then watch GPU and CPU load, since a miner shows up in utilization graphs first.
Black Lotus Labs' own advice is more basic still: audit what is exposed every time a new open source tool is installed, since in this campaign the distance from default configuration to infected server was a single crafted request. The operator needed only to change four words to redirect every bot, and did so 11 times. The defenders' equivalent is just as small, a firewall rule on a port that should never have been open.