Security researchers have uncovered a malware campaign named PoeLLMthat has already compromised over 3,400 internet-facing serverssince April 2026. Its objective: hijack exposed artificial intelligence (AI) and large language model (LLM) infrastructure to secretly mine cryptocurrency. The campaign is financially motivated and shows new levels of stealth and reach.
What PoeLLM Does and How It Spreads
PoeLLM infiltrates servers through known vulnerabilities in enterprise tools like LiteLLM, Gotenberg, Gitea, and Ivanti Sentry appliances when they’re exposed to the internet. Once inside, it installs crypto miners (including XMRig and another called Iron) and connects them to a service known as Kryptex based in Russia. Compromised machines aren’t just used for mining—they also become scouts, scanning for additional vulnerable systems and delivering further payloads.
One of the campaign’s more inventive features is how it encodes its command-and-control (C2) server addresses. Attackers host a poem in a public GitHub repository and embed the server address in particular words and “keys” inside that poem. Every time they set up a new C2, they tweak a few words to shift the derived address—a kind of poetic obfuscation.
Scope, Geography & Attribution
Since its activity began in April 2026, PoeLLM has emerged from relative obscurity to impact more than 3,400 servers. Mid-June saw a peak with nearly 2,200 servers compromised and around 800 turning active daily. The infections are primarily located in the United States and Western Europe.
Cyber threat analysts attribute the operation to an Italian-speaking actor, supported by Italian-language code artifacts and traffic patterns observed across compromised hosts.
Why AI/LLM Infrastructure Is a Prize Target
Exposed AI and LLM services provide especially attractive attack surfaces. Not only do they often run on powerful hardware—ideal for cryptocurrency mining—but they also may contain valuable data. The attackers behind PoeLLM are exploiting this double value: gain computation power for mining, and possibly access to sensitive model or operational data.
Moreover, servers compromised by PoeLLM don’t just perform mining. They are turned into exploit launchpads—probing for further exposed instances and using HTTP POST requests to coax new victims into pulling down payloads from the malware’s C2 servers.
At present, the attacker’s brute-force login capabilities (e.g. via SSH) show signs of experimentation—traffic toward SSH portals and other login endpoints indicate attempts to expand through credential-based attacks. But this functionality appears to still be under development.
Defending against PoeLLM requires both patching known vulnerabilities in exposed AI/LLM systems and auditing externally reachable services to ensure nothing is accidentally open. Monitoring outbound connections could also help detect devices reaching back to the poem-based C2 infrastructure.
Analytical Take:PoeLLM underscores how attackers are pivoting toward the AI/LLM stack—not just its models, but the entire hardware-and-infrastructure footprint. As AI deployments scale, especially within enterprises seeking to leverage LLMs in production, the exposure risk increases dramatically. This campaign blends cunning obfuscation (poem-based C2), high-value computing use, and opportunistic compromise of internet-faced systems. Moving forward, AI infrastructure owners must treat attack surface management and network hygiene as mission-critical: patch early, minimize exposure, and assume the compute itself will be a target.