Research published November 18; IPCRIME briefing November 19, 2026. A Rapid7 study examined 20,354 preselected IPv4 addresses on November 10. After filtering likely deception infrastructure, it found unauthenticated AI or notebook APIs on 7,627 hosts. About 55% of its analyzed Ollama inventories contained model-name artifacts suggesting outside activity.
Those indicators included marker names, external registry references and encoded file paths. They do not independently prove successful exploitation or data theft. The study used read-only requests, without inference, writes or exploitation. Its selected sample cannot establish internet-wide prevalence, and a listed cloud model does not prove that an account was used or billed.
IPCRIME analysis: compare the inventory with an accountable owner
A useful first review links each model to a deployment request, an owner and an intended workload. Keep the observed list before changing it. Ask whether an unexpected entry corresponds to an approved experiment, an administrator’s test or activity nobody can explain. The answer should come from change records and available logs, not from whether the model’s name sounds reputable.
Next, record who can reach the service and who pays for any connected account. These may be different people. A researcher who installed a runtime may not know that a later deployment exposed it through a public gateway. Establish one owner for the complete route from client to model service.
Use the documented local controls
Ollama’s local FAQ says the default listener is 127.0.0.1:11434. It also documents disabling cloud features through configuration or OLLAMA_NO_CLOUD=1, followed by a restart. Those options let an operator align a local-only workload with its intended design.
Keep exposure and incident conclusions separate
If a service should be private, close the unintended access path while preserving relevant records. Review linked accounts and usage separately, and document whether credentials were actually available to the exposed system. Escalate unexplained activity through the incident process without claiming that a model-list entry alone proves a stolen file.
For the next review, define an expected inventory and a named person who approves changes. That creates a concrete comparison point and makes it easier to distinguish an authorized experiment from an unexplained addition.
Related reading: Digital Arsenal: Audit Vendor Keys and Exposed Secrets With Open-Source Tools.
Featured image: conceptual IPCRIME artwork, not a photograph or screenshot of this incident or product.
