Dispatch, 2026-11-12, SHC Team, AI security
Thousands of exposed servers, most running AI tools, were taken over
Filed 2026-11-12 · SHC Team
On Oct. 7, researchers at Lumen's Black Lotus Labs reported that a malware campaign had infected thousands of servers since April, mostly in the United States and Western Europe. Most victims appeared to be running open-source AI tools, such as LiteLLM and Ollama, on servers anyone on the internet could reach.
What happened
Lumen calls the malware PoeLLM. It carries cryptocurrency-mining software and gives the attacker a remote shell, a way to type commands on an infected server. It then turns infected servers into scanners that hunt for the next victims. The researchers say AI servers draw attackers because they may hold useful data and run on powerful hardware that suits cryptocurrency mining.
Lumen says the attackers got in by exploiting flaws in services open to the public internet. For LiteLLM, Lumen says the likely way in was a flaw tracked as CVE-2026-42271 (a CVE is the public ID number for a security flaw). By April 21, LiteLLM's developer, BerriAI, had published a security notice pointing to a fixed version. The flaw let anyone holding even a low-level access key run their own commands on the server.
On June 8, CISA, the federal Cybersecurity and Infrastructure Security Agency, added that flaw to its Known Exploited Vulnerabilities list, which records flaws it has evidence attackers are using. On Sept. 2, CISA added another LiteLLM flaw to the list, this one in how the software checks logins.
What changes for you
Running AI on your own server changes who holds your data. It does not, on its own, keep attackers out. Lumen describes these AI servers as "self-hosted, internet-exposed services with known vulnerabilities." What hurt them was being reachable from the internet while running out-of-date software. In Lumen's words: "Delays in updating internet-facing AI and enterprise tools enable highly trusted access."
Lumen also notes that new tools "often go unmonitored for vulnerabilities." For a small office, law firm, farm co-op or nonprofit, that could be an AI trial someone set up and then forgot. A server taken over this way works for the attacker, who can also reach the data on it.
What you can do
- Ask whether it can be reached from the internet. If an AI tool only serves your team, ask your IT person or provider whether it needs to be open to the internet at all. Lumen advises checking what is exposed after installing any new open-source tool.
- Ask who installs security fixes, and how fast. Every AI tool needs a named person or provider who updates it. Ask whether they watch CISA's free Known Exploited Vulnerabilities list.
- Ask about keys and logins. Change any default keys, give each person their own, and remove keys nobody uses. In the LiteLLM flaw, even a low-level key was enough to run commands on the server.
- Turn off what nobody uses. Ask whether anyone set up an AI tool to test it, and shut down the ones you no longer need.
- Or have someone else run it. SHC can set up private AI for your team, then watch and fix it on terms set in your contract. Security is a goal we share with clients, not a promise, so ask us these same questions.
Sources
- Lumen Black Lotus Labs, "Canto incognito: tracking the PoeLLM malware", 2026-10-07
- BerriAI (LiteLLM) GitHub security advisory, "Authenticated command execution via MCP stdio test endpoints" (CVE-2026-42271), 2026-04-21
- CISA, "CISA Adds Two Known Exploited Vulnerabilities to Catalog", 2026-06-08
- CISA, "CISA Adds Seven Known Exploited Vulnerabilities to Catalog", 2026-09-02