Skip to content

Security1 publisher2 min readPublished

6,935 exposed Ollama servers answered an internet scan without asking for credentials

Mysterium VPN counted 36,769 self-hosted AI endpoints reachable from the public internet. Ollama is the only population where the scan can prove that nothing on the host asked for credentials.

The Watch · Security desk

Photograph accompanying 6,935 exposed Ollama servers answered an internet scan without asking for credentials
Photo: sentinelone.com

What happened

  • Mysterium VPN queried the Netlas scanning index and counted 36,769 reachable self-hosted AI endpoints across model servers, agent-building platforms and vector stores.
  • Only 2.02% of those endpoints returned an HTTP authentication challenge, which the report describes as no network-layer gate whatsoever for the overwhelming majority.
  • Open WebUI was the largest single population at 18,529 reachable instances, and exactly one of them returned an authentication challenge.
  • The Ollama population came to 6,935 hosts, of which 6,046 returned an explicit HTTP 200 to the scanner.
  • Agent builders and workflow platforms accounted for 5,223 of the total, among them Flowise, n8n, ComfyUI, Dify, RAGFlow, Langflow and Open WebUI Pipelines.

Compiled by The WatchSomething wrong?How this is made

Why it matters

  • capability An outsider who reaches one of these Ollama endpoints can enumerate the models installed on the machine and spend the owner's GPU time generating text, with the owner paying the power bill. The resource abuse is known as LLMjacking.
  • exposure A workflow platform holds whatever secrets its automations need: OpenAI API keys, database credentials, Slack tokens, webhook secrets, CRM passwords.
  • contradiction SentinelOne and Censys put exposed Ollama hosts at around 175,000 in January, roughly 25 times Mysterium's figure, from a different index and a looser fingerprint.
  • constraint The 401-or-403 test cannot tell an operator whether a particular 200 is an open API or a login form, so no one can read their own risk off the 2.02% without testing their own deployment.

Mysterium calls 36,769 a lower bound. The method explains why. The researchers treated an HTTP 401 or 403 as evidence of an authentication gate and refused to read a 200 as the opposite, because a login page returns 200 as readily as an open API does [5]. They queried Netlas, matched fingerprints against the responses and stopped there [7]. "We counted the doors. We didn't open them," the report said [6]. Nobody connected to a host, ran a model or read a credential, so nothing in this study shows that any of these endpoints was abused [7].

Ollama is the exception because of how it answers. Its root endpoint returns the string "Ollama is running" to a request carrying no credentials, so the fingerprint that identifies the service is also the proof that the service did not challenge the caller [11].

2.02% of 36,769 is about 743 endpoints with a gate in front of them [1]. Open WebUI supplied one of those, and vLLM's 4,880 endpoints supplied three [8][9], which leaves roughly 739 gated hosts somewhere other than the two largest populations in the study [3]. Open WebUI on its own is 50.4% of the census [7]. LocalAI's 150 instances and llama.cpp's 69 returned no challenge at all [10].

The workflow layer is where the credentials sit. Mysterium counted 1,341 reachable Flowise instances and not one HTTP authentication challenge among them [20]. The report also points at CVE-2026-40933, a critical Flowise flaw that lets an authenticated attacker run arbitrary commands through the MCP adapter, and which Flowise has fixed [21]. Authentication is a real precondition on an instance that asks for it, and costs the attacker nothing on an instance that never asks.

The exclusions are large. Mysterium recorded 22,024 responses on Ollama's default port 11434 and left all of them out of the total, because a port number does not prove which service is listening on it [17]. That excluded set is about 3.2 times the size of the Ollama population the report was willing to count [4].

What to watch

  • A follow-up that authenticates, or fails to, against the Open WebUI instances returning 200 would settle whether 18,529 is an exposure number or a login-page number.
  • Reports of CVE-2026-40933 being exploited against unpatched public Flowise instances with no authentication configured.
  • Whether a re-scan on Netlas moves the Ollama count toward the SentinelOne and Censys figure or away from it.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories