Skip to navigation

Threat reportThreat IntelligenceTL-2026-6345

Exposed Unauthenticated Ollama/AI Servers Marked and Probed by Third Parties (Rapid7: 'Open by default. Already occupied.')

highACTIVE

Rapid7's 10 Nov 2026 scan of 20,354 pre-selected IPv4 hosts found 7,627 hosts answering AI/notebook APIs without authentication, almost all Ollama. 54.8% of the 7,675 exposed Ollama servers carried third-party model names/tags (callback registries, tunnel pulls, marker names, encoded file-path payloads) indicating intent to probe or abuse; successful exploitation is not confirmed.

Key facts

IdentifiersTL-2026-6345, CVE-2024-37032
Also known asOpen by default. Already occupied., Probllama (related CVE-2024-37032), Self-hosted LLMjacking (hypothesis)
SeverityHIGH
First disclosed
Exploitation statusPublic proof-of-concept exploit for a linked CVE
EPSSCVE-2024-37032: 89.6% (99.7th percentile)
Affected productsOllama Ollama; Project Jupyter Jupyter Notebook / JupyterHub; LF AI & Data / MLflow MLflow Tracking Server; Streamlit / Grafana Labs Streamlit and Grafana (presence checks only)
Weaknesses (CWE)CWE-22, CWE-306, CWE-918
Malwaresolarmarker
ToolsOllama, Rapid7
MITRE ATT&CK techniquesT1071.001 Application Layer Protocol: Web Protocols, T1082 System Information Discovery, T1098.004 Account Manipulation: SSH Authorized Keys, T1190 Exploit Public-Facing Application, T1543.002 Create or Modify System Process: Systemd Service, T1552.001 Unsecured Credentials: Credentials In Files, T1552.004 Unsecured Credentials: Private Keys, T1552.005 Unsecured Credentials: Cloud Instance Metadata API, T1572 Protocol Tunneling, T1574.006 Hijack Execution Flow: Dynamic Linker Hijacking, T1595 Active Scanning
MotivationUNKNOWN
Target sectorstechnology, cloud-hosting, research-and-academia, any-organization-self-hosting-ai
Target regionsEurope, Asia, North America
Detection rules9 (1 shown free)
Sources9 cited sources

Detection coverage

As of 2026-11-18, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-6345 across Splunk SPL, Microsoft KQL and Sigma, covering 20 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: Ollama API Abuse - Marker Models, Callback Registries and Encoded Path-Traversal Payloads (CVE-2024-37032 pattern)

Detects HTTP POSTs to the unauthenticated Ollama API (TCP 11434) on /api/pull, /api/push, /api/create, /api/copy or /api/delete whose request body carries the indicators Rapid7 documented for TL-2026-6345: callback registry IPs (24.144.109.68:9882, 103.249.201.108:9882, 76.13.49.213:9882, 62.31.247.107:9882, 43.143.113.68:8899, 185.165.171.21:8899), free tunnel or ephemeral registries (trycloudflare.com, ttl.sh, bore.pub, Pinggy), marker model names (zzvitals-beacon, verif_sys, ops-verify, dsv4probe, pwn_ctl_), Base64-encoded file paths (/etc/shadow, /proc/self/environ, /etc/passwd, /etc/ld.so.preload, authorized_keys) and ../ digest traversal as used in the Probllama chain (CVE-2024-37032). Hits show attempted probing or marking of the server; successful exploitation is not implied.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1190, T1082, T1552.001, T1574.006, T1098.004
  • Data sources: splunk_stream_http, network_traffic, ollama_api_11434
  • Severity: high
  • Confidence: high

Splunk SPL

index=network sourcetype="stream:http" dest_port=11434 http_method=POST uri_path IN ("/api/pull","/api/push","/api/create","/api/copy","/api/delete")
| eval body=coalesce(form_data, http_request_body, "") ```base64 path prefixes are case-sensitive so the body is not lowercased```
| eval registry_callback=if(match(body,"(?i)((24[.]144[.]109[.]68|103[.]249[.]201[.]108|76[.]13[.]49[.]213|62[.]31[.]247[.]107):9882|(43[.]143[.]113[.]68|185[.]165[.]171[.]21):8899|159[.]223[.]110[.]159)"),1,0)
| eval tunnel_registry=if(match(body,"(?i)(trycloudflare[.]com|ttl[.]sh|bore[.]pub|pinggy)"),1,0)
| eval marker_model=if(match(body,"(?i)(zzvitals-beacon|verif_sys|ops-verify|dsv4probe|pwn_ctl_)"),1,0)
| eval encoded_path=if(match(body,"(L2V0Yy9zaGFkb3c|L3Byb2Mvc2VsZi9lbnZpcm9u|L2V0Yy9wYXNzd2Q|L2V0Yy9sZC5zby5wcmVsb2Fk|L3Jvb3QvLnNzaC9hdXRob3JpemVkX2tleXM)"),1,0)
| eval digest_traversal=if(match(body,"(?i)([.][.]/|%2e%2e)"),1,0) ```Probllama style digest path traversal```
| eval signal_count=registry_callback+tunnel_registry+marker_model+encoded_path+digest_traversal
| where signal_count>0
| eval src_zone=if(cidrmatch("10.0.0.0/8",src_ip) OR cidrmatch("172.16.0.0/12",src_ip) OR cidrmatch("192.168.0.0/16",src_ip),"internal","external")
| stats count as requests min(_time) as firstTime max(_time) as lastTime values(uri_path) as endpoints sum(registry_callback) as registry_callback_hits sum(tunnel_registry) as tunnel_registry_hits sum(marker_model) as marker_model_hits sum(encoded_path) as encoded_path_hits sum(digest_traversal) as traversal_hits max(signal_count) as max_signals by src_ip src_zone dest_ip
| convert ctime(firstTime) ctime(lastTime)
| sort - max_signals

Microsoft KQL

let IocTerms = dynamic(['ttl.sh','bore.pub','trycloudflare.com','pinggy','zzvitals-beacon','verif_sys','ops-verify','dsv4probe','pwn_ctl_','24.144.109.68','103.249.201.108','76.13.49.213','62.31.247.107','43.143.113.68','185.165.171.21','159.223.110.159']);
Syslog
| where TimeGenerated > ago(24h)
| where ProcessName =~ 'ollama' or SyslogMessage contains '[GIN]'
| where SyslogMessage has '/api/' or SyslogMessage has_any (IocTerms)
| extend ClientIP = extract(@'[|] *([0-9a-fA-F.:]+) *[|] *(?:POST|GET|DELETE)', 1, SyslogMessage)
| extend Endpoint = extract(@'(?:POST|GET|DELETE) +.(/api/[a-z]+)', 1, SyslogMessage)
| extend IocHit = SyslogMessage has_any (IocTerms)
| extend IsExternal = isnotempty(ClientIP) and ipv4_is_private(ClientIP) == false and ClientIP !in ('127.0.0.1','::1')
| summarize StateChangeCalls = countif(IsExternal and Endpoint in ('/api/pull','/api/push','/api/create','/api/copy','/api/delete')), ReconCalls = countif(IsExternal and Endpoint in ('/api/tags','/api/ps','/api/show','/api/version')), IocHits = countif(IocHit), Endpoints = make_set(Endpoint, 10), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated) by Computer, ClientIP
| where StateChangeCalls > 0 or IocHits > 0 or ReconCalls >= 10
| project FirstSeen, LastSeen, Computer, ClientIP, StateChangeCalls, ReconCalls, IocHits, Endpoints
| order by IocHits desc, StateChangeCalls desc

Sigma

title: Ollama State-Changing API Call or Path Traversal Against Exposed Ollama Server
id: 7c1e4b2a-93d5-4f0e-8a67-2b5d1c9e3f41
status: experimental
description: Detects POST requests to the Ollama pull, push, create, copy or delete API endpoints, or Ollama API requests containing directory traversal sequences, as seen when exposed unauthenticated Ollama servers are marked or probed (TL-2026-6345, Rapid7 research) and in the CVE-2024-37032 Probllama exploit chain.
references:
  - https://www.rapid7.com/blog/post/tr-self-hosted-ai-renting-compute-exposed-servers
  - https://www.wiz.io/blog/probllama-ollama-vulnerability-cve-2024-37032
  - https://nvd.nist.gov/vuln/detail/CVE-2024-37032
author: Threadlinqs Team
date: 2026-11-18
tags:
  - attack.initial-access
  - attack.t1190
  - cve.2024-37032
logsource:
  category: webserver
detection:
  selection_state_change:
    cs-method: 'POST'
    c-uri|contains:
      - '/api/pull'
      - '/api/push'
      - '/api/create'
      - '/api/copy'
      - '/api/delete'
  selection_api:
    c-uri|contains: '/api/'
  selection_traversal:
    c-uri|contains:
      - '../'
      - '..%2f'
      - '%2e%2e'
  condition: selection_state_change or (selection_api and selection_traversal)
falsepositives:
  - Administrators or CI pipelines legitimately pulling or creating models through a reverse proxy
  - Authorized internal AI platform tooling such as Open WebUI back ends
level: medium

Known false positives:

  • Administrators legitimately pulling or copying models from internal or approved registries over the Ollama API
  • Internal vulnerability scanners or Rapid7-style exposure scans that send Ollama API probes

Tuning notes: Requires Splunk Stream (or equivalent) capturing request bodies on port 11434; map form_data to your body field. Add known admin source IPs and approved registry hostnames to an exclusion lookup. Treat src_zone=external with any signal as high priority; the IP list is Rapid7-observed and may include research scanners or sensors, so confirm before blocking. Rapid7 could not confirm any target ran a vulnerable build, so triage host version (fixed in 0.1.34) alongside the alert.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Ollama External State-Changing and Recon API Calls Plus IOC Strings in Server Logs (Sentinel Syslog) · unlock with TL Creds
  • Ollama Process Writes ld.so.preload, Systemd Unit or authorized_keys (Defender for Endpoint on Linux) · unlock with TL Creds
  • Ollama Egress to Callback Registries, Tunnel Services and Cloud Metadata (Defender for Endpoint) · unlock with TL Creds
  • Ollama Model Pull From Raw IP Registry or Free Tunnel Service (SSRF-Style Callback) · unlock with TL Creds
  • Command-Line Access to Ollama Private Key or Service Override Config (Credential Theft) · unlock with TL Creds
  • Ollama Sign-In or Cloud Model Use on Server (Self-Hosted LLMjacking Precursor, Hypothesis) · unlock with TL Creds
  • Ollama Process Touching Persistence and Credential Paths (ld.so.preload, systemd, authorized_keys, shadow, environ) via auditd · unlock with TL Creds
  • Ollama Callback Registry IPs and Tunnel Registry Pulls (trycloudflare, [indicator], [indicator], Pinggy) · unlock with TL Creds

Indicators of compromise

As of 2026-11-18, Threadlinqs Intelligence tracks 20 indicators of compromise for TL-2026-6345.

Sample indicators (3 of 20)

  • IP address: 24.144.109[.]68
  • Domain: ttl[.]sh
  • IP address: 103.249.201[.]108

Defanged, TLP:CLEAR. Full set of 20 indicators: unlock with TL Creds.

20 indicators of compromise · unlock with TL Creds. Compare plans

Community OSINT corroboration

1 of this threat's indicators have also been reported by the open-source security community. Community sightings are unverified and are kept separate from Threadlinqs' curated indicators. Indicator values, reporters and campaign linkage are available to authenticated Red-tier users.

Technical analysis

Rapid7 (Christiaan Beek) scanned a pre-selected set of 20,354 IPv4 addresses on 10 November 2026 across ports 3000, 5000, 8000, 8080, 8501, 8888 and 11434, sending one unauthenticated HTTP request per service (Ollama /api/tags, Jupyter /api/kernels, MLflow experiment search, JupyterHub /hub/api/, Streamlit and Grafana health endpoints). No inference, write or exploitation requests were sent. Raw scanner output was 14,064 findings on 11,891 IPs; after removing 1,513 findings on 429 likely honeypot/sensor hosts (148 hosts sharing an identical five-service llama3.1:8b fingerprint and 281 hosts where one IP:port answered as multiple products), critical findings fell from 387 to 39. 7,714 findings on 7,627 hosts were confirmed unauthenticated AI or notebook APIs, almost all Ollama (7,675 servers listing 106,938 model entries and 18,733 distinct tags). Only 0.3% were served over HTTPS and about 65% sit in hosting/cloud networks (OVHcloud, Hetzner, Contabo, Oracle Cloud, Tencent Cloud, DigitalOcean account for 42%). Top countries: Germany 1,306, China 1,130, United States 1,097, France 549, India 301.

The central observation is that 54.8% (4,209) of exposed Ollama servers held artifacts left by third parties, visible only as model names/tags. 40.3% (3,091) carried marker or 'I was here' models (e.g. zzvitals-beacon:latest, verif_sys:latest, ops-verify:latest, dsv4probe:24h, pwn_ctl_<host-ip>_ae52...:latest); 34.7% (2,666) had models pulled from a non-official registry, including 23.7% from a raw IP:port and 22.4% via a free tunnel or temporary registry (trycloudflare.com, ttl.sh, bore.pub, Pinggy); 15.5% (1,190) held gpt-4o / gpt-4 / claude-3-opus aliases usually alongside tinyllama; 712 model names embed the IP of the host they sit on. The most frequent callback registries were 24.144.109.68:9882 (trigger/benign, trigger/plamo2-<id>), 103.249.201.108:9882, 76.13.49.213:9882, 62.31.247.107:9882, 43.143.113.68:8899 (library/tens) and 185.165.171.21:8899 (recon/probe). Rapid7 describes the pattern as Find -> List -> Call back -> Mark -> Reach, i.e. SSRF-style registry pulls used for egress confirmation and inventory tagging. Marker timestamps run from 10 June to 3 November 2026, and some June markers were still present on 10 November, suggesting nobody owns or monitors those servers.

55 hosts (0.7%) carried model names embedding Base64/hex-encoded file paths (e.g. L2V0Yy9zaGFkb3c = /etc/shadow, L3Byb2Mvc2VsZi9lbnZpcm9u = /proc/self/environ) that resemble path-traversal probes: host fingerprinting (47 hosts, /etc/passwd, /proc/self/environ), secrets in environment or service configuration (28 hosts, Ollama systemd unit and override.conf), credentials and keys (15 hosts: /etc/shadow, SSH authorized_keys, AWS config, the Ollama private key on 10 hosts), persistence paths (11 hosts: /etc/ld.so.preload, new systemd service files) and cloud metadata (1 host, Oracle Cloud). These mirror the exploit chain of CVE-2024-37032 ('Probllama', Wiz Research): a missing digest-format validation in Ollama before 0.1.34 permits path traversal via /api/pull (arbitrary file write), /api/push (file read) and code execution by corrupting /etc/ld.so.preload. Rapid7 cannot confirm whether any target ran a vulnerable build, so the artifacts show attempted intent rather than success.

A second risk is 'self-hosted LLMjacking' (Rapid7 hypothesis, untested): 79% of exposed Ollama hosts list at least one :cloud model (GLM, DeepSeek, MiniMax, Kimi variants) that routes to ollama.com when the host is signed in via 'ollama signin', which links the local ed25519 key at ~/.ollama/id_ed25519 to the owner's account; the same seven cloud models appear on 1,295 unrelated hosts, suggesting automated checks. 28% of hosts run abliterated/heretic/uncensored models. No threat actor is named or attributed; the callback registries may reflect large-scale research scanning, and some remaining 'critical' hosts may still be sensors. Third-party indicators come only from model names, so a clean list does not prove a clean host.

MITRE ATT&CK techniques by tactic:

Command and Control

T1071.001 Application Layer Protocol: Web Protocols; T1572 Protocol Tunneling

Discovery

T1082 System Information Discovery

Persistence

T1098.004 Account Manipulation: SSH Authorized Keys; T1543.002 Create or Modify System Process: Systemd Service

Initial Access

T1190 Exploit Public-Facing Application

Credential Access

T1552.001 Unsecured Credentials: Credentials In Files; T1552.004 Unsecured Credentials: Private Keys; T1552.005 Unsecured Credentials: Cloud Instance Metadata API

stealth

T1574.006 Hijack Execution Flow: Dynamic Linker Hijacking

Reconnaissance

T1595 Active Scanning

Remediation

Patches

  • Upgrade Ollama to 0.1.34 or later (CVE-2024-37032 fix)

Immediate actions

  • Scan owned and expensed cloud/VPS ranges (including non-standard ports) for Ollama (11434), Jupyter, MLflow, Streamlit and Grafana exposed without authentication
  • Review every Ollama model list (/api/tags) for IP-prefixed names, trycloudflare/ttl.sh/bore.pub/Pinggy pulls and marker names such as zzvitals-beacon, verif_sys, ops-verify, dsv4probe, pwn_ctl_*
  • Treat any hit as compromise: rotate secrets, revoke the ollama.com key linked to ~/.ollama/id_ed25519, check cloud spend, and rebuild the host
  • Block egress to the callback registries 24.144.109.68, 103.249.201.108, 76.13.49.213, 62.31.247.107, 43.143.113.68 and 185.165.171.21

Workarounds

  • Place Ollama behind a reverse proxy enforcing authentication (Wiz recommendation)
  • Firewall port 11434 and other AI/notebook ports from the public internet

Longer-term hardening

  • Bind AI runtimes to localhost or private networks and front any remote access with an authenticating reverse proxy
  • Restrict egress from model servers to approved registries only
  • Log and alert on state-changing Ollama API calls (pull, create, copy, push, delete)
  • Require a named owner, data classification and expiry for every model server and add AI runtimes to exposure management and red-team scope
  • Avoid running 'ollama signin' on shared servers without authentication and spend alerts

Affected and fixed versions

  • Ollama — Ollama
    Vulnerable versions: Any instance exposed without authentication on 11434; All versions before 0.1.34 (CVE-2024-37032)
    Fixed in: 0.1.34 (path-traversal fix only; authentication must be added separately)
  • Project Jupyter — Jupyter Notebook / JupyterHub
    Vulnerable versions: Instances answering /api/kernels or /hub/api/ without authentication
  • LF AI & Data / MLflow — MLflow Tracking Server
    Vulnerable versions: Instances answering experiment search without authentication
  • Streamlit / Grafana Labs — Streamlit and Grafana (presence checks only)
    Vulnerable versions: Instances exposed to the internet; Grafana default-credential labels were inferred, not tested

Timeline

  • Wiz Research disclosed the Ollama digest path-traversal flaw (later CVE-2024-37032, 'Probllama') to the Ollama maintainers.
  • Ollama 0.1.34 released with digest validation fix (sources give 7 or 8 May 2024).
  • CVE-2024-37032 published in NVD (CVSS 3.1 8.8 HIGH, CWE-22; affects Ollama before 0.1.34).
  • Wiz publicly detailed Probllama; reporting noted many internet-exposed Ollama instances were still running vulnerable versions.
  • Earliest third-party marker-model timestamp observed on exposed Ollama servers; some June markers were still present on 10 November.
  • Latest embedded marker timestamp observed in exposed Ollama model names; most timestamps cluster in October 2026.
  • Rapid7 scanned 20,354 pre-selected IPv4 hosts and found 7,627 hosts with unauthenticated AI/notebook APIs; 54.8% of exposed Ollama servers carried third-party artifacts.
  • Rapid7 published 'Open by default. Already occupied.' documenting callback registries, marker models and encoded file-path probes.

Sources cited

Other reports on CVE-2024-37032

Drafted by the Threadlinqs research pipeline from 9 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats