AI Governance Institute
← News
Research2026-10-07

PoeLLM Malware Hits 3,000+ Servers by Hiding Commands Inside AI-Read Poetry

What happened

Lumen Black Lotus Labs published research on a campaign it named Canto Incognito, documenting how malware called PoeLLM has infected more than 3,000 enterprise servers since April 2026. The targets were servers running open-source AI inference tools, specifically LiteLLM and Ollama, which organizations use to run AI models on their own hardware rather than through cloud providers. Attackers attributed to an Italian-speaking actor embedded instructions telling the malware where to report back, hiding those instructions inside a poem posted to GitHub. When the AI system fetched and read that poem, the poetic phrasing bypassed the model's safety filters, allowing the hidden commands to pass through undetected. This is the first confirmed real-world use of adversarial poetry as a method to evade AI safety guardrails. Infected servers were used for cryptomining, where attackers use another organization's computing resources to generate cryptocurrency, and were also recruited into botnets that searched for additional vulnerable AI infrastructure.

Why it matters

  • ·AI inference servers running open-source tools like LiteLLM and Ollama are now confirmed attack targets at scale. Organizations that self-host AI infrastructure for data-residency or cost reasons, common in regulated industries, must treat these endpoints as governed, security-reviewed IT assets rather than developer utilities.
  • ·The adversarial poetry technique confirms that AI safety guardrails can be bypassed by embedding harmful instructions in content the model reads from external sources. This is a form of indirect prompt injection, where an attacker plants instructions in a document or webpage that an AI system later retrieves and acts on. Compliance teams should verify that adversarial robustness testing covers externally fetched content, not only direct user inputs.
  • ·The campaign exposed a governance ownership gap. AI infrastructure components often fall between the AI governance team, which focuses on model behavior, and the IT security team, which may not have inventoried self-hosted AI tools at all. Enterprises without a complete inventory of AI serving infrastructure cannot assess or remediate this exposure.

Stay ahead of stories like this

Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.

Free every Thursday. Unsubscribe anytime.

Governance controls affected

What to do now

  • ☐Ask your engineering and IT security teams to produce a complete list of every server or container running open-source AI inference tools, including LiteLLM, Ollama, and similar products, and confirm whether any of those endpoints are reachable from outside your internal network.
  • ☐Verify that your adversarial testing program covers scenarios where an AI system retrieves content from external sources, such as web pages, documents, or code repositories, not only scenarios involving direct user inputs.
  • ☐Confirm that your AI infrastructure servers are covered by the same vulnerability scanning, patching cadence, and access controls applied to other production systems, and escalate any gaps to your CISO.
  • ☐Review whether your AI incident response playbook covers infrastructure-layer compromises, such as a model-serving server being taken over for cryptomining, and assign clear ownership for this category of incident.
  • ☐If your organization self-hosts AI inference tools for data-residency or cost reasons, brief your AI governance committee on the Canto Incognito campaign and request a formal risk assessment of your self-hosted AI attack surface.

What to watch next

Security researchers and threat intelligence teams are likely to document additional campaigns targeting open-source AI infrastructure as PoeLLM-style techniques become known to other threat actors. Compliance teams should monitor whether CISA or sector-specific regulators issue advisories on AI inference endpoint security, particularly for critical infrastructure operators. Hiding instructions inside AI-readable content on trusted platforms like GitHub has supply chain implications. Organizations using AI tools to fetch external content in automated workflows should expect updated guidance from framework bodies. The Five Eyes Guidance on the Careful Adoption of Agentic AI Services already identifies external content retrieval as a risk vector. Further regulatory attention to this class of attack is likely.

Related Coverage

Standards2026-11-16

Microsoft Defense Guide Makes Layered Prompt Injection Controls an Enterprise Baseline

Microsoft has published updated guidance on defending against indirect prompt injection, a category of attack where malicious instructions hidden in external content redirect AI systems to perform unintended actions. The guidance, titled 'Defend against indirect prompt injection attacks,' recommends a layered set of controls including prompt shields, content isolation, behavioral monitoring, and human approval gates. It applies to any organization using AI systems that process third-party content such as emails, documents, websites, or plugins.

Research2026-11-14

One Rogue Employee-Built AI Agent Exposed Five Identity and Control Gaps

Token Security published a case study tracing how a single AI agent created five governance failures. The study draws on a Cloud Security Alliance survey of 418 IT and security professionals. The failures span agent identity, access provisioning, prompt injection, overprivileged static credentials, and behavior-based detection. The findings apply to any enterprise where employees can build agents using no-code or AI-assisted tools.

Corporate Policy2026-11-11

Bay's $50M Agent Control Platform Exposes the Runtime Oversight Gap

Tel Aviv-based Bay emerged from stealth on November 11, 2026, with $50 million in funding to build a runtime control layer for enterprise AI agents. The platform automatically inventories agents, tools, and connections, then evaluates every proposed agent action in real time before allowing, redirecting, or blocking it. The launch reflects a recognized enterprise gap: most organizations cannot see what their AI agents are doing or stop them from taking unintended actions.