PoeLLM Malware Leverages GitHub Poem Keywords for C2 Rotation Targeting Exposed AI Services

PoeLLM has updated its guiding poem eleven times with an Italian-speaking operator maintaining the infrastructure. Infected hosts parse the poem for keywords that map directly to octets, enabling rapid C2 failover without hard-coded domains or IPs that signature-based tools would catch. The technique targets exposed AI and open-source services rather than traditional enterprise endpoints.

The SecurityWeek roundup pairs this with South Korean authorities investigating bank intrusions where President Lee Jae Myung cited signs of AI involvement in customer data theft. No technical indicators from the Korean incidents have been released, creating an attribution gap between observed creative C2 methods and official suspicion of AI tooling.

Independent verification of PoeLLM remains limited to Black Lotus Labs telemetry; no public CVE or exploit chain has been tied to the Korean breaches. The pattern shows operators repurposing public AI infrastructure for persistence and monetization while governments flag AI use without disclosing matching artifacts.

Next steps include monitoring GitHub for further poem revisions and Korean police release of IOCs that could link the two activity sets. Agencies should prioritize behavioral detection on exposed AI endpoints over keyword or domain blacklists.