Reference
Common AI/LLM Infrastructure Ports
Source: https://github.com/nuclide-research/AI-LLM-Infrastructure-OSINT/blob/main/reference/ports
Cross-cutting port reference for AI/ML-adjacent services. Useful for Shodan/Censys/FOFA queries, nmap/naabu scan prioritization, firewall audits, and asset discovery.
Sorted numerically. Where a port hosts multiple AI/ML services, the primary ones are listed first.
Quick-Reference Table
| Port | Service(s) | Notes |
|---|---|---|
| 80 / 443 | Generic HTTP(S), Dify, Flowise, reverse-proxied everything | Filter by http.title: / HTML fingerprint |
| 1337 | Jan, Devika | Hacker-cute defaults |
| 1984 | LangSmith | |
| 2375 | Docker daemon (unauth) | RCE → host foothold |
| 2379 | etcd (Milvus metadata), Kubernetes control plane | |
| 3000 | Flowise, Open WebUI, AnythingLLM, AgentGPT, SuperAGI, Langfuse, Promptfoo, OpenDevin, Grafana | Most crowded port in AI |
| 3001 | AnythingLLM | |
| 4000 | LiteLLM Proxy | Provider keys live here |
| 4040 | Apache Spark UI | Often co-deployed with ML pipelines |
| 4317 | OpenTelemetry gRPC (OTLP) | LLM observability transport |
| 4318 | OpenTelemetry HTTP (OTLP) | LLM observability transport |
| 4567 | Rivet | |
| 5000 | MLflow | Models, artifacts, experiments |
| 5001 | KoboldCpp | |
| 5050 | pgAdmin | Often default creds |
| 5432 | PostgreSQL + pgvector, Supabase, Neon, Timescale | |
| 5500 | ChromaDB (alt) | |
| 5601 | Kibana, OpenSearch Dashboards | Vector index admin |
| 5678 | n8n | AI workflow automation |
| 6006 | Phoenix/Arize, TensorBoard | Traces + training viz |
| 6333 | Qdrant (HTTP) | Snapshots downloadable |
| 6334 | Qdrant (gRPC) | |
| 6379 | Redis / Redis Stack (vector search) | Often no auth |
| 6443 | Kubernetes API server | ML workload orchestration |
| 6900 | Argilla | RLHF/annotation data |
| 7474 | Neo4j Browser | Graph memory stores |
| 7501 | Lightning AI | |
| 7687 | Neo4j Bolt, Memgraph | |
| 7860 | Gradio, LangFlow, unsloth, text-generation-webui | HuggingFace Spaces default |
| 7997 | Infinity (embeddings) | |
| 8000 | LangChain, vLLM, Triton, FastAPI generic, ChromaDB, AutoGPT, BentoML, Ray Serve, MetaGPT, Mem0, many /v1/* OpenAI-compat | Single most common LLM port |
| 8001 | RedisInsight | |
| 8008 | ClearML | |
| 8080 | LocalAI, llama.cpp, Vespa, BabyAGI, Axolotl, Determined AI, Kubeflow, Airflow, Helicone, Dgraph, NVIDIA, Vast.ai, HF TEI/TGI, Phidata | Generic “alt-HTTP” |
| 8081 | mongo-express | |
| 8088 | Hadoop YARN ResourceManager | Training data pipelines |
| 8089 | Splunk HEC | Sometimes LLM log sink |
| 8108 | Typesense | API key enumeration risk |
| 8123 | LangGraph Studio, ClickHouse | |
| 8161 | ActiveMQ Web Console | ML pipeline message broker |
| 8265 | Ray Dashboard | Cluster job submission, RCE |
| 8443 | SageMaker Notebook, alt-HTTPS | |
| 8501 | Streamlit | |
| 8529 | ArangoDB | |
| 8787 | Cloudflare AI Gateway, Portkey, RStudio Server | |
| 8882 | Marqo | |
| 8888 | Jupyter, RunPod | RCE if no token |
| 9000 | MinIO (Milvus backing), Portainer | Vector blobs in buckets |
| 9090 | Prometheus | Every ML stack exports metrics here |
| 9091 | Milvus metrics, Zilliz | |
| 9092 | Apache Kafka | LLM event streams, training pipelines |
| 9200 | Elasticsearch / OpenSearch | dense_vector / kNN |
| 9400 | NVIDIA DCGM | GPU telemetry |
| 9870 | Hadoop NameNode (HDFS) | Training data at rest |
| 9998 | Apache Tika | Document ingestion |
| 10250 | Kubelet | K8s node attack surface |
| 11434 | Ollama | Most-exposed LLM runtime in 2025-26 |
| 19530 | Milvus (gRPC) | |
| 27017 | MongoDB | Increasingly used as vector store |
| 50070 | Hadoop NameNode (legacy) |
Patterns Worth Knowing
Highest-yield single-port queries
port:11434, catches tens of thousands of unauth Ollama instances. Model-naming leaks org context (e.g.acme-internal-rag:latestdiscloses tenant identity before authenticating).port:7860, Gradio/HuggingFace Spaces ecosystem; favicon hash-1294819032catches reverse-proxied instances that strip the title.port:8000+"/v1/chat/completions", OpenAI-compatible endpoint regardless of underlying engine. vLLM, LM Studio, llama.cpp, LocalAI, text-generation-webui all collapse into this.port:6006, historically TensorBoard, now also Phoenix/Arize for LLM traces. Same port, very different exposure surface: training metrics vs. live prompt/response logs.
Port ranges by deployment pattern
- Enthusiast / self-hosted: 1337, 4567, 7501, 8123, 11434, less likely to have org-grade auth.
- Production-ish web UIs: 3000, 8080, 7860, SaaS-flavored, sometimes wrapped in Cloudflare but often not.
- Data plane (never meant to be public): 5432, 6379, 9200, 19530, 27017, direct database exposure = bypass of any auth layer the app imposes.
- Control / admin planes: 2375, 2379, 6443, 8265, 9000, one of these exposed means the entire workload is owned.
Co-location signals
If you see these ports together on one host, infer the stack:
11434 + 3000→ Ollama + Open WebUI (self-hosted private ChatGPT)6333 + 8000 + 4000→ Qdrant + a RAG app + LiteLLM proxy5000 + 8265 + 9400→ MLflow + Ray + NVIDIA DCGM (training cluster)9200 + 5601 + 3000→ Elasticsearch + Kibana + Langfuse (RAG observability stack)8888 + 8265→ Jupyter + Ray (notebook-driven ML cluster)
Ports that are always worth a second look
| Port | Why |
|---|---|
| 2375 | Unauth Docker daemon = host RCE via /containers/create |
| 8265 | Ray cluster job submission = arbitrary code execution across GPU nodes |
| 6443 / 10250 | K8s API / Kubelet = full workload control if unauth |
| 11434 | Ollama exposes /api/generate → free inference + model enumeration |
| 4000 | LiteLLM proxy admin UI leaks master provider keys if misconfigured |
Scanning Tips
# Fast first-pass of common AI ports with naabu
naabu -host <target> -p 2375,3000,4000,5000,5432,6006,6333,6379,7860,8000,8080,8265,8888,9090,9200,11434,19530
# Nmap service detection on the long list
nmap -sV -p 80,443,1337,1984,2375,2379,3000,3001,4000,4040,4317,4318,4567,5000,5001,5050,5432,5500,5601,5678,6006,6333,6334,6379,6443,6900,7474,7501,7687,7860,7997,8000,8001,8008,8080,8081,8088,8089,8108,8123,8161,8265,8443,8501,8529,8787,8882,8888,9000,9090,9091,9092,9200,9400,9870,9998,10250,11434,19530,27017,50070 <target>
# Shodan facet to see which ports dominate a given org
shodan search --facets port org:"Example Inc" "api" | headContributing
New AI-adjacent port? Open a PR adding a row (keep sorted numerically) with the service name and a Notes column entry when the exposure has non-obvious impact. See CONTRIBUTING.md.