Ollama Exposure Leaves AI Servers Open to Compute Abuse

Thousands of internet-facing AI servers are exposing unauthenticated APIs and showing signs that someone else has already used them, according to a Rapid7 snapshot published Nov. 13. The research identified 7,714 findings across 7,627 hosts where an AI or notebook API answered without authentication, almost all involving Ollama, the open-source runtime for running large language models.

Rapid7 scanned 20,354 preselected IPv4 addresses on seven common service ports on Nov. 10, using read-only HTTP requests to detect Ollama, Jupyter, MLflow, Streamlit, Grafana and JupyterHub. After removing likely honeypots and deceptive catch-all services, 7,675 Ollama hosts remained. The researchers caution that the sample is not an internet-wide census and that exposure alone does not prove compromise; only 0.3% of all findings were served over HTTPS.

The more serious finding was evidence of third-party activity in the model inventories returned by exposed Ollama APIs. Rapid7 said 54.8% of the remaining Ollama hosts contained markers such as models pulled from raw IP addresses, temporary tunnel services or suspicious names containing terms including beacon, verify and probe. On 55 hosts, encoded model names decoded to targets including /etc/shadow, environment and service configuration files, SSH authorization files, cloud metadata and Ollama’s private key. These indicators suggest reconnaissance, callback testing or attempts to establish persistence, but the scan did not send inference, write or exploitation requests and did not establish that those attempts succeeded.

Cloud-account abuse is another risk. Rapid7 found that 79% of exposed Ollama hosts listed at least one cloud-routed model. Ollama’s hybrid mode can route requests through the local daemon to Ollama’s cloud service when a user has signed in, linking an account key stored under ~/.ollama/id_ed25519. An exposed, signed-in server could therefore become an unauthorized gateway to a customer’s paid model subscription, with usage charges and possible data exposure falling on the owner. The researchers did not test inference or claim that every listed cloud model represented an active account.

Organizations should also verify the runtime version. Ollama versions before 0.1.34 are affected by CVE-2024-37032, a path-traversal flaw in digest handling that NVD rates 8.8 High under CVSS 3.1; the issue can enable file writes and, in vulnerable deployments, remote code execution. Ollama fixed it in version 0.1.34, and Wiz Research documented the vulnerability as Probllama. Rapid7 did not attribute the observed activity to a named threat actor, identify confirmed exploitation in the scan, or claim that exposed hosts were still running the vulnerable version.

Defenders should inventory public cloud and VPS assets, bind Ollama and notebook services to localhost or private networks, and place any required remote access behind authentication or a zero-trust gateway. Teams should review model lists and API logs for unfamiliar pulls, creates, copies and pushes; restrict outbound connections to approved registries; rotate credentials and cloud keys on systems showing suspicious artifacts; check for unexpected model-service spending; and rebuild suspected-compromised hosts rather than treating the incident as a simple configuration error. Cloud resource abuse has long been used for cryptomining, but the Rapid7 findings show how exposed AI runtimes can also turn compute, model access and linked cloud accounts into rentable infrastructure for strangers.

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply