Most recent
navigate open esc close Corpus index built 2026-06-07 23:58 UTC

← All reference

Reference

1. LLM Orchestration Platforms

Source: https://github.com/nuclide-research/AI-LLM-Infrastructure-OSINT/blob/main/shodan/queries/01-llm-orchestration

Section verified: April 30, 2026

Low-code/no-code builders, agent runtimes, and chain orchestrators. These platforms typically expose a web UI that, when unauthenticated, grants direct access to flow editors, API keys stored in nodes, and execution endpoints.

Flowise

Shodan QueryNotes
title:"Flowise" port:443586 hits, title fingerprint, mixed auth
product:"Flowise"576 hits, Shodan product facet, canonical fingerprint
http.html:"Low-code LLM apps builder"572 hits, HTML title fingerprint
"Flowise"170 hits, broad banner match
"Flowise" http.status:20026 hits, 200-response subset, likely reachable
"Flowise" "chatflows"9 hits, banner + API term
"Flowise" "Express"6 hits, Express-fronted subset
"X-Powered-By: Express" "Flowise"6 hits, header + banner intersect
"Flowise" http.status:4014 hits, auth-enabled subset

Deployment note: As of April 2026, port:3000 (Flowise default) is effectively dead on the public internet, production deployments now sit behind 443 reverse proxies almost universally. Prefer product:"Flowise" as the canonical fingerprint.

CVE note: CVE-2024-36420 (auth bypass via path traversal) affects Flowise < 1.8.2. Exposed pre-auth instances are RCE candidates, not just info disclosure.

Other Orchestrators

Shodan QueryNotes
http.html:"open-webui"19,549 hits (2026-04-30), package/asset string in HTML body; broader than title-only match, catches builds where the title was customized
http.title:"Open WebUI"18,736 hits, largest AI UI fingerprint on the internet; Ollama frontend
http.title:"Open WebUI" port:80802,842 hits (2026-04-30), direct default-port subset; first-user-admin = effectively unauth, treat as T1
http.html:"dify"8,750 hits, broad Dify HTML fingerprint
http.title:"LiteLLM"5,076 hits, LLM proxy, master key often leaked in env
"Jan" port:13374,624 hits, desktop app in server mode
http.title:"Dify"2,614 hits, tighter Dify title fingerprint
http.title:"Clawdbot Control"1,770 hits ⚠️ http.title: is tokenized, sample before trusting, may include false positives
http.html:"Chainlit"1,144 hits, conversational UI layer on LangChain
http.title:"Langflow"844 hits, flow builder, often deployed unauth
"AnythingLLM" port:3001366 hits, known auth bypass history
http.title:"Gradio"225 hits, generic Gradio wrapper (covers oobabooga, demos, custom AI apps)
port:18789 ("openclaw" OR "clawdbot")165 hits, OpenClaw gateway (grouped OR required; unparenthesized breaks Shodan precedence)
"LocalAI" port:808095 hits, no auth by default
http.html:"Ollama is running" -port:44326,580 hits (2026-04-30), canonical root-response string; -port:443 drops TLS-fronted noise; T1, Ollama has no auth support, exposure = full model access
product:Ollama port:1143421,067 hits (2026-04-30), Shodan product facet on default port; direct-exposure subset, T1
"Ollama" port:1143437 hits, banner-only fallback; the two queries above are the canonical fingerprints
"hexstrike-ai" port:11434Ollama with HexStrike AI offensive orchestrator loaded, 47 MCP tools, unrestricted shell exec when Flask server running
"qwen3.5-abliterated" port:11434Refusal-stripped 35B MoE Qwen model, no safety guardrails
"huihui_ai" port:11434huihui-ai abliterated model family (qwen3/qwen3.5 variants)
"abliterated" port:11434Broad match for refusal-removed models on exposed Ollama instances
http.html:"AutoGPT"32 hits, project moribund since 2025, retained for completeness
http.favicon.hash:-140453829311 hits, LlamaIndex favicon
"LangChain" port:80006 hits, library fingerprint, app varies
http.title:"Create Llama App"6 hits, LlamaIndex default UI (RAG starter)
http.html:"haystack"3,341 hits, ⚠️ generic term, collides with search tools, monitoring, GDS frameworks
"zylon"73 hits, PrivateGPT vendor; ⚠️ collides with anime/gaming names
http.html:"privategpt"7 hits, PrivateGPT-specific HTML match

Verified April 2026. Deployment note: the “service + default port” pattern that dominated 2024 is largely dead, most platforms moved behind 443/80 reverse proxies. Queries below 10 hits are retained when they still identify the platform uniquely.

OpenClaw / Clawdbot: This is not a passive reconnaissance target. A publicly reachable OpenClaw gateway is an agent with shell execution, browser automation, email send, and calendar write on whoever deployed it. Treat positive hits as live compromise surface, not data disclosure.

Prompt / Chain Management

Shodan QueryNotes
product:"n8n"77,102 hits, canonical n8n fingerprint; RCE history (CVE-2024-25289 and successors), see n8n note below
"n8n"4,966 hits, banner-only, narrower subset
http.title:"n8n"370 hits (2026-04-30), title-level match, often editor UI; basicauth optional and frequently skipped, lean T1
http.html:"/rest/login"162 hits (2026-04-30), n8n REST login path leaked into HTML body; tighter than title-only when present
"x-powered-by: Express" port:5678134 hits (2026-04-30), Express header on n8n’s default port; direct-exposure subset bypassing reverse-proxy fronting
http.title:"n8n - Workflow Automation"24 hits (2026-04-30), verbose default title; high-confidence editor UI subset
http.html:"langgraph"501 hits, LangGraph Studio / LangChain graph orchestrator
http.html:"rivet"169 hits, ⚠️ polluted; “Rivet” collides with Rivet Networks NIC UIs, storage products
http.title:"Rivet"71 hits, ⚠️ same pollution concern
http.title:"LangGraph"51 hits, title-level match
http.html:"promptflow"5 hits, best PromptFlow variant found
http.title:"PromptFlow"4 hits, title match

n8n note: n8n is by far the most-exposed workflow/orchestration platform observed in this catalogue, roughly 4× the count of Open WebUI and ~130× Flowise. The default-port fingerprint (port:5678) is obsolete as of April 2026; nearly all deployments sit behind reverse proxies, n8n.cloud, or containerized ingress. Prefer product:"n8n" as the canonical query. Given n8n’s “execute code” and HTTP-request nodes, exposed editors with weak or default auth are direct RCE surface, not just workflow disclosure.

FP traps (verified 2026-06-04, Cat-01 dogfood run #4)

DorkTrapEvidence
http.title:"Langflow"~100% FP. Tokenized title; Shodan reported 96,070 “available” vs catalog’s 844. 0 of 16 sampled fingerprinted as real Langflow.Insight #15 at the extreme - a tokenized title dork can be near-total FP, not just ~50%. Use http.html:"langflow" or a body marker instead.

Honeypot fleet caught (Cat-01): an 8-IP MCP “credential-theft” bait fleet (identical tool surface get_aws_admin_credentials/get_ssh_session_credentials/ schedule_commands, every IP Shodan-tagged honeypot, aimap-profile score 8) surfaced on ports 3001/443. aimap reported 13 false criticals. Now refuted by VisorCAS honeypot-host-tagged screen signature (commit 0f60fac).


Virgin re-birth dork set — 2026-06-04

Sourced from the Stage -1 6-squad pre-assessment (data/platform-intel/llm-orchestration-osint-2026-06-04.md). Hit counts are NOT filled in — run via Playwright (feedback_shodan_playwright_only), log each to shodan/query-log.md. FP-risk is the design-time estimate; the marker-probe (see intel doc §3) is what promotes a hit to a finding.

Builders (Squad Alpha)

DorkPlatformFP-riskBasis
http.title:"Flowise - Build AI Agents" port:3000FlowiseLowfull multi-token title from index.html
port:3000 "flowData" "chatflowid"FlowiseLowAPI response fields
port:7860 "langflow_version"LangflowLow/api/v1/config field; replaces FP title dork
port:7860 "chat" "db" "status" "ok"LangflowLow-Medhealth_check triple-field shape
port:80 "step" "setup_at"DifyMedsetup endpoint body; exclude github
http.title:"Langflow"Langflow~100% FPtokenized title — DO NOT USE (FP-trap table above)

Western front-ends (Squad Bravo)

DorkPlatformFP-riskBasis
http.favicon.hash:1470014414Open WebUILowverified mmh3 hash
http.title:"Open WebUI"Open WebUILow-MedSSR title
http.title:"LibreChat"LibreChatLow-MedSSR HTML shell
http.title:"LobeChat"LobeChatLow-MedNext.js SSR metadata (verify)
http.title:"big-AGI"big-AGILow-MedSSR title (unverified)
http.title:"Chatbot UI"Chatbot UIHighgeneric string — needs supabase conjunct

Chinese-OSS (Squad Charlie)

DorkPlatformFP-riskBasis
http.html:"ragflow" http.html:"RAGFlow"RAGFlowLowboth case forms
http.html:"infini_rag_flow"RAGFlowVery lowdefault service pw leaks in configs/errors
http.title:"Coze Studio" port:8888CozeLowByteDance-exclusive
http.html:"opencoze" port:8888CozeVery lowDB name in bundle paths
http.html:"fastgpt-" port:3000FastGPTLowAPI key prefix in bundle refs
http.html:"dataelement" port:3001BISHENGLowvendor asset path

Frameworks / visual builders (Squad Delta)

DorkPlatformFP-riskBasis
http.html:"chainlit-cloud.s3.eu-west-3.amazonaws.com"ChainlitLowframework-injected og:image
http.html:"github.com/Chainlit/chainlit"ChainlitLowdefault og:url
port:23333 http.html:"promptflow"PromptFlow svcLowheartbeat body anchor
http.html:"Botpress is loading"BotpressLowinit-state HTML, version-stable
port:21888RivetHighnon-distinctive port; likely null-result

Workflow + agent gateways (Squad Echo)

DorkPlatformFP-riskBasis
http.title:"n8n.io - Workflow Automation"n8nLowexact title
port:5678 "REST_ENDPOINT" http.status:200n8nLowJS constant, aimap-verified
port:18789 "Clawdbot Control"OpenClawLowconfirmed live-scan dork
port:18789 http.html:"clawdbot-app"OpenClawLowReact root id
port:1865 "cheshire-cat-logo"Cheshire CatLowlogo asset
port:42110 "/server/admin"KhojLow-MedDjango admin on non-std port

Local runtimes (Squad Foxtrot)

DorkPlatformFP-riskBasis
"Ollama is running"OllamaVery lowexact root string
port:11434OllamaLownear-exclusive port
port:4891GPT4AllLowunusual port; localhost-default → small pop
port:8080 "/system" "backends"LocalAILowbackends array exclusive
port:7860 "TextGen"oobaboogaMedtitle; add /v1/internal/* probe
port:7860 "h2oGPT"h2oGPTMedbody branding; /openai_api/v1/ prefix is the low-FP discriminator

Zero-result protocol: 0 hits → vary the signature (favicon, body marker, alt port, API-path angle), do not conclude absent (METHODOLOGY §3). Log the zero to query-log.md.