ChromaDB on Tier-2 Cloud: Auth Posture Survey (Scope Expansion)

NuClide Research · 2026-05-04 Companion to: chromadb-cloud-survey-2026-05.md (DO/Hetzner/Vultr baseline, 48 instances) Sibling tier-2 expansions: ollama-tier2-cloud-survey-2026-05.md, qdrant-tier2-cloud-survey-2026-05.md, milvus-tier2-cloud-survey-2026-05.md

Summary

Mass-scan of port 8000 (ChromaDB default) across the same 76 tier-2 /16 ranges (3.55M IPs), Scaleway + OVH + Linode used in the parallel Qdrant/Milvus/Ollama tier-2 expansions. 34,524 port-open candidates → 44 confirmed ChromaDB instances → 23 populated.

DCWF KSAT coverage

Auto-derived from DCWF AI work-role rule files (ksat-tag).

• 672 (AI Test & Evaluation Specialist): K7003, S7068, S7075, T5858, T5904
• 733 (AI Risk & Ethics Specialist): K7040, K7051, T5854, T5868, T5904
• overlap (Common AI KSATs (all 5 roles)): K1157, K1158, K22, K6935, K7003, K942

Combined with the original DO/Hetzner/Vultr baseline (48 instances), total ChromaDB count across NuClide surveys is 92 instances, 100% unauthenticated, the auth-off-default thesis reproduces cleanly.

The notable findings are not the auth state itself (ChromaDB ships auth-off; that’s now well-documented) but the branded enterprise tenant content appearing in unauth ChromaDB collections at population scale:

• STIHL (German power-tools brand) AI deployment via RaptorCX integrator, host serves ai-stihl.raptorcx.com cert
• AXA Insurance, rag_axa collection on unauth ChromaDB
• Mitsubishi + Daikin, branded Japanese corporate tenants
• Indonesian financial regulator (OJK) + Indonesian Personal Data Protection Law (UU PDP), RAG corpus over Indonesian banking regulations and the privacy law itself
• Healthcare/medical content, oncology, patient_info_embeddings, larvol_kol (Larvol pharma KOL data)

Methodology

masscan -iL <76 tier-2 /16 CIDRs> -p 8000 --rate 10000
  → 34,524 port-8000 hits (port 8000 is heavily shared)

chroma-probe.py (200-thread, strict signature)
  GET /api/v1/heartbeat → JSON containing "nanosecond heartbeat"
  GET /api/v1/version
  GET /api/v1/collections OR /api/v2/tenants/default_tenant/databases/default_database/collections
  → 44 confirmed ChromaDB instances

Strict heartbeat-signature probe filters out the 99.9% non-ChromaDB hits on port 8000 (port also used by uvicorn FastAPI apps, Open WebUI, generic HTTP servers).

Findings Summary

Combined ChromaDB across all NuClide surveys

Operator identification (cert pivots on populated tier-2 hosts)

13 of 23 populated hosts have TLS certs revealing operator identity (identities redacted pending coordinated-disclosure windows):

Remaining 10 populated hosts have no TLS cert on port 443 (OVH default rDNS) and are not directly operator-attributable.

Notable collection content

Sample of distinctive collection names observed across the 23 populated hosts:

The mix of major-brand tenants (AXA, Mitsubishi, Daikin), regulatory/government content (OJK, UU PDP, Hilversum municipality), and healthcare/pharma content (oncology, patient_info_embeddings, larvol_kol) on shared unauth ChromaDB instances is the same pattern observed in the parallel Qdrant tier-2 survey: operators don’t recognize that ChromaDB on port 8000 is a separate security perimeter from their authenticated front-end, and so the data tier carrying enterprise/regulated content is exposed.

Cross-platform correlation

• Two-tier auth-skew pattern (also documented in Qdrant tier-2): Multiple ChromaDB tier-2 hosts run a Sermon Live-style chat app on port 80/443/3000 (auth-protected) while exposing the backing ChromaDB on port 8000 unauth. Same operator pattern: front-end auth perimeter, data-tier accidental exposure.
• Brand-tenant overlap with Milvus tier-2: Just as the Milvus tier-2 survey found OCBC Bank Singapore + STIHL-class branded tenants on unauth Milvus, the ChromaDB tier-2 surface includes AXA + STIHL + Mitsubishi + Daikin. Major brands are commissioning AI-RAG products from integrators who are using auth-off-default vector DBs as backing stores. The integrators bear remediation responsibility but the brands face reputational exposure.
• Government / regulatory content: OIF Francophonie (Qdrant tier-2), Quebec municipal-government RAG (Milvus tier-2), Indonesian OJK + UU PDP (ChromaDB tier-2), Hilversum Dutch municipality (ChromaDB tier-2). At the population scale, government/regulatory content is meaningfully represented on auth-off-default vector stores.

Disclosure posture

Per the consolidated ledger at disclosure/qdrant-snapshot-disclosure-ledger-2026-05.md, no per-operator disclosures have been drafted for the ChromaDB tier-2 cohort yet, the Qdrant disclosures are in the active 30-day window and the cumulative outreach is being managed at sustainable cadence. Per-operator drafts can be added on demand for ChromaDB hosts with the highest content-sensitivity (oncology / patient-info / branded enterprise tenants), but the aggregate finding is the same auth-off-default thesis.

The recommended fix is the same as for the broader survey series:

1. Set CHROMA_SERVER_AUTHN_PROVIDER and CHROMA_SERVER_AUTHN_CREDENTIALS (token or basic auth)
2. Bind ChromaDB to 127.0.0.1 and access only over a private network
3. Apply identical posture to the front-end’s data backend, not just the front-end itself

Raw Data

~/recon/chromadb-tier2-2026-05-04/tier2-chroma-confirmed.jsonl  (44 hosts)

See Also

• chromadb-cloud-survey-2026-05.md, DO/Hetzner/Vultr baseline (48 instances)
• qdrant-tier2-cloud-survey-2026-05.md, sibling Qdrant tier-2 (663 unauth)
• milvus-tier2-cloud-survey-2026-05.md, sibling Milvus tier-2 (36 real, 393 honeypot pollution)
• SYNTHESIS-2026-05.md, cross-survey synthesis paper