Specialty Data Layers — Shodan Query Catalog

Generated: 2026-05-27 from pre-survey OSINT pass (15 platforms) See: data/platform-intel/specialty-data-layers-osint-2026-05-27.md for full intel

Platforms covered: ClickHouse, Apache Cassandra, Redis Stack, MinIO, Feast, Hopsworks, Tecton, Feathr, ArangoDB, Neo4j, Apache Kafka REST Proxy, Apache Flink, Spark History Server, Trino/Presto, Delta Sharing Server.

---

ClickHouse

Auth default: off (default user ships with empty password; pre-DeepSeek-incident Docker images leave the default user network-accessible with no password) Exposure class: Full SQL access to all databases; system tables expose env vars, query logs, schema, LLM request/response logs stored in AI stacks. SELECT * FROM system.environment dumps environment variables including any secrets.

---

Apache Cassandra

Auth default: off (authenticator: AllowAllAuthenticator is the yaml default; no credentials required) Exposure class: Full CQL read/write to all keyspaces; JMX on 7199 gives full cluster management. AI stacks expose feature vectors, session data, time-series ML features.

---

Redis Stack / Redis with Vector Modules

Auth default: off (no requirepass by default; ~60,000 of 300,000+ internet-exposed instances have zero auth per Wiz 2025) Exposure class: Full key-value read/write including all embeddings, cached API keys, LLM conversation history, feature values. CVE-2025-49844 (RediShell, CVSS 9.9) enables RCE via Lua on no-auth instances.

---

MinIO

Auth default: default-creds (minioadmin:minioadmin Docker default; CVE-2023-28432 leaked root credentials via unauthenticated POST on pre-2023-03-20 releases) Exposure class: Full S3 API access to all buckets — model weights, training datasets, MLflow artifacts, DVC cache, Kubeflow pipeline artifacts. CVE-2023-28432 (CISA KEV) leaks MINIO_SECRET_KEY and MINIO_ROOT_PASSWORD in plaintext.

---

Feast

Auth default: off (auth: type: no_auth is the documented default in feature_store.yaml; TLS also off by default) Exposure class: Full online feature store read/write — real-time feature values for any entity, RAG document embeddings via /retrieve-online-documents, feature metadata revealing data schema. /push endpoint enables data poisoning.

---

Hopsworks

Auth default: default-creds (admin@kth.se / admin documented default from official installer and community forums) Exposure class: Full ML platform access — feature groups, feature views, training datasets, model registry (model binaries + metadata), experiment runs. Complete ML artifact lifecycle exposed.

---

Tecton

Auth default: on (managed SaaS, API-key required; no open self-hosted deployment model) Exposure class: N/A — no open-auth exposure documented.

Note: Expected near-zero actionable hits. Any findings are likely customer portals or documentation sites, not feature store infrastructure.

---

Feathr

Auth default: off for Docker sandbox (local development image; Azure AD required for production) Exposure class: Feature registry read/write — feature definitions, transformation logic (Python UDFs), entity definitions. Business logic exposure even without raw data access.

---

ArangoDB

Auth default: off (--server.authentication defaults to false; Docker image warns but does not enforce auth) Exposure class: Full multi-model database access — graph data (knowledge graphs, RAG entity maps), document collections, key-value store. Foxx microservice framework enables RCE if app services are deployed. /_api/database/user lists all databases without auth.

---

Neo4j

Auth default: default-creds (neo4j/neo4j, forced password change on first login); many Docker deployments set NEO4J_AUTH=none for development and expose it externally Exposure class: Full Cypher query access — knowledge graph entities and relationships, RAG document chunk metadata, ontology definitions. In AI stacks: LLM-extracted entity graphs, concept maps, user behavior graphs for recommendation.

---

Apache Kafka REST Proxy

Auth default: off (REST Proxy default config binds to port 8082 with no auth; “the REST Proxy bypasses broker ACLs when authorization is disabled”) Exposure class: Full topic enumeration — message stream names reveal business process topology. Consumer group access enables message replay. In AI pipelines: real-time feature events, inference requests, training data streams. Native Kafka on 9092 (no auth by default) allows producing/consuming all messages.

---

Apache Flink

Auth default: off (“the REST endpoint does not authenticate the client by default” per official Flink docs) Exposure class: Running job visibility, full cluster config (may include Kafka brokers, DB connection strings, AWS credentials via /jobmanager/config), JAR upload enabling RCE. CVE-2020-17518/17519 (arbitrary file write/read) affect unpatched instances <= 1.11.2.

---

Spark History Server

Auth default: off (“security features like authentication are not enabled by default in Apache Spark”) Exposure class: Spark job history including environment variables from all historical jobs — AWS access keys, S3 bucket names, Hive metastore passwords, Databricks PATs stored as Spark config properties. Job names reveal pipeline structure.

---

Trino / Presto

Auth default: off (“Trino runs with no security by default”; port 8080, no auth, no TLS in default config) Exposure class: Full SQL query submission to any connected data source — S3/Delta Lake/Iceberg/Hive. Running query text and results exposed via /v1/query without auth. Cluster topology via /v1/cluster.

---

Delta Sharing Server

Auth default: bearer-token (required by protocol; exposure risk is static/demo tokens with zero expiry, or tokens embedded in documentation used in production) Exposure class: With valid bearer token: all shared Delta Lake tables — training datasets, feature tables, model evaluation sets. Token misconfiguration (zero expiry, demo tokens in production) is the primary risk.