
Exposed Self-Hosted AI Infrastructure at Internet Scale
Security researcher Christiaan Beek has published new internet-wide telemetry on the Rapid7 Blog examining how self-hosted artificial intelligence services are being deployed without authentication. Drawing on a snapshot of 20,354 internet-facing hosts, the November 18, 2026 report identifies 7,627 systems exposing confirmed unauthenticated AI or notebook APIs. Rapid7 compares the pattern to the wave of open databases left online a decade ago, noting that many of these instances appear to operate on rented compute rather than inside traditional corporate data centers.
Rather than documenting purely theoretical misconfigurations, the dataset captures widespread evidence of prior external interaction. According to Rapid7, 55 percent of exposed Ollama servers already carry model artifacts left behind by someone other than the machine owner. Even more strikingly, 79 percent of the exposed Ollama instances list cloud-routed models, turning unauthenticated inference ports into open gateways backed by someone else's paid upstream API account.
Architecture of the Exposure: Port 11434 and Model Tag Artifacts
At the center of the technical findings is Ollama's default HTTP interface on port 11434. An unauthenticated, read-only GET request to the /api/tags endpoint returns the server's list of registered models. Because automated scanners and opportunistic actors repeatedly probe open inference ports, Beek observes that those model listings effectively function as a persistent visitors' log across thousands of machines.
Sample output redacted and defanged by Rapid7 illustrates three distinct categories of entries coexisting on exposed servers. First are standard local text, coding, and embedding models such as llama3.1:8b, qwen3.6:27b, qwen2.5-coder:7b, gemma4:e4b, tinyllama:latest, smollm2:135m, nomic-embed-text:latest, bge-m3:latest, and qwen3-embedding:0.6b, alongside tags labeled gpt-4o:latest, claude-3-opus:latest, and gpt-4o:latest-copy-test. Second are cloud-routed and uncensored entries, including glm-5.3:cloud, deepseek-v4-pro:cloud, kimi-k2.7-code:cloud, minimax-m3:cloud, and mistral-large-4:cloud, as well as abliterated builds like huihui_ai/Qwen3.8-abliterated:latest and Huihui-CyberStrike-OffSec-35B-abliterated.
The third category consists of third-party reconnaissance and callback strings embedded directly into model names. Observed tags include beacon and verification identifiers such as zzvitals-beacon:latest, verif_sys:latest, ops-verify:latest, and ttl[.]sh/dsv40110/dsv4probe:24h, an IP-stamped control marker formatted as pwn_ctl_<this-host-ip>_ae52..., and direct callback endpoints like 103.249.201[.]108:9882/trigger/benign:latest. Rapid7 also documented Cloudflare tunnel addresses under trycloudflare[.]com carrying Base64-encoded Linux paths, including L2V0Yy9zaGFkb3c—which decodes to /etc/shadow—and L3Byb2Mvc2VsZi9lbnZpcm9u, decoding to /proc/self/environ.
Impact on Open-Source AI Workflows and Cloud Accounts
In practical developer workflows, these exposures frequently arise when teams deploy lightweight local inference servers or notebook environments on temporary cloud virtual machines to evaluate coding assistants, embedding pipelines, or frontier model proxies. Because Ollama's unauthenticated API is convenient for automated scripts and local tools, leaving port 11434 exposed to the public internet turns short-lived test infrastructure into shared compute for uninvited external traffic.
For open-source practitioners spinning up self-hosted inference stacks on leased GPU servers, the findings underscore the operational hazard of open-by-default configurations. Services designed for rapid local experimentation on a developer workstation become immediately reachable to automated internet scanners when bound to public network interfaces without firewall rules, VPN gating, or reverse-proxy authentication.
The combination of local weights, abliterated models, and :cloud tags exposes operators to layered abuse beyond unauthorized local inference. When 79 percent of exposed Ollama servers list cloud-routed models, any external party capable of reaching the API can potentially route requests through the owner's metered cloud credentials. Meanwhile, the presence of offensive-security model variants and callback-style tag strings demonstrates that outside actors actively catalog, test, and tag exposed hosts rather than merely passively scanning port 11434.
Honeypot Filtering, Methodology, and Evidence Boundaries
Rapid7 pairs its headline exposure counts with explicit warnings about scanner noise and attribution limits. Although the initial scan flagged 14,064 total findings and 387 critical issues, filtering out likely honeypots reduced the critical count by nearly an order of magnitude, dropping from 387 down to 39 confirmed critical findings. Highlighting that gap upfront signals how easily raw internet scan tallies can overstate real-world impact when decoy infrastructure is left unfiltered.
To keep observed telemetry distinct from broader conclusions, Beek labels claims across the report as [Observed], [Interpretation], or [Hypothesis]. The core dataset relies on passive, read-only queries to endpoints like /api/tags across 20,354 hosts rather than intrusive host-level forensics. As a result, while the artifact logs establish widespread third-party contact on 55 percent of exposed Ollama nodes, they do not independently prove full host compromise on every tagged machine, nor do they detail individual cloud provider policies, model licensing terms, or regional compliance obligations for the affected deployments.
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-19.



