The vector database
that stays encrypted even while it's searching.

Similarity search directly on encrypted data — self-hosted in your VPC or on-prem. Only the answer to a query is decrypted.

Get Started Request Demo

<5ms

p95 encrypted query latency

100M+

vectors per index

Zero

data leaves your boundary

Your storage never touches plaintext.

CyborgDB runs entirely in your environment — Cyborg holds no keys, sees no vectors, and never receives your data. Only the final candidates are ever decrypted.

Embed & encrypt

Your model produces a vector. CyborgDB encrypts it with your key, within your own VPC or on-prem.

AES-256-GCM · in your VPC

Search on ciphertext

A trapdoor query runs against the forward-secure encrypted index. The backing store only ever holds ciphertext — on disk, in snapshots, and on the wire.

Encrypted IVF / HNSW

Rerank & return

Only the final candidates are decrypted to rerank, then results return over TLS. The per-operation key is zeroized after.

transient · zeroized

Everything a production vector DB should do.
None of the security tradeoffs.

Encrypted ANN

Approximate nearest neighbor on ciphertext, with recall within 2% of plaintext baselines.

· IVF + HNSW backends

· k up to 1,000

· cosine, dot, L2

Metadata filters

Combine vector similarity with encrypted metadata predicates in one request. Hybrid keyword search is on the roadmap.

· AND/OR filter trees

· Equality + range on encrypted tokens

◐ Hybrid keyword — coming soon

Per-field keys

Different fields can use different keys. Revoke access to a field without re-indexing the rest.

· Independent key rotation

· Field-level audit log

· Partial revocation

Hot & cold tiers

Move indexes between NVMe and object storage with a single API call. Both tiers stay fully encrypted.

· S3 / GCS / Azure Blob

· Manual promote / demote

· Cold tier @ $0.001/GB/mo

BYOK + HSM

Bring your own key from AWS KMS, GCP KMS, Azure Key Vault, or an on-prem HSM.

· PKCS#11 HSM

· Envelope encryption

· Auto key rotation

Snapshots

Encrypted snapshots to your bucket on demand or on a schedule. Restore in-place or to a new index.

· Object-lock + WORM compatible

· Cross-region replication

· Restore in < 10 min

A vector DB API you already know.

Five lines to index and query. The service encrypts and indexes your vectors inside your environment, and searches encrypted vectors.

client.py

from cyborgdb import Client

client = Client(base_url="http://localhost:8000",

api_key="your-api-key")

# generate or load an encryption key

index_key = client.generate_key(save=True)

index = client.create_index(index_name="my_index", index_key=index_key)

# vectors sent over TLS, encrypted server-side

index.upsert([

{"id": "item_1", "vector": [0.1, 0.2, 0.3, 0.4], "contents": "Hello!"},

])

# search runs on ciphertext

results = index.query(query_vectors=[0.1, 0.2, 0.3, 0.4], top_k=10)

Designed for a strong adversary —
without holding your data.

CyborgDB is self-hosted; we never receive your vectors, payloads, or keys. The encryption model means an attacker with disk access to your cluster gets ciphertext only.

Defense in depth

Six guarantees that hold even if an attacker owns your disk, snapshots, and network.

Encrypted search Embeddings stay encrypted through storage, transit, and index traversal. Only the final candidates are decrypted.
Forward privacy New inserts reveal nothing about prior data; current queries can't be correlated with historical patterns.
Inversion-resistant Encrypted embeddings defeat the ML reconstruction attacks that recover 99% of plaintext from standard vector DBs.
Key isolation Encryption keys live in your KMS, HSM, or local store. CyborgDB operates with zero knowledge of them.
Per-record randomization Unique IVs per record block cross-system metadata correlation, even with full disk access.
AEAD on the wire TLS plus AES-256-GCM at the application layer. Endpoint compromise doesn't downgrade transport security.

The only residual exposure is the service process mid-query; it's stateless and zeroizes on completion. Compliance frameworks (SOC 2, HIPAA, GDPR, FedRAMP) apply to your environment, not ours — CyborgDB is a binary you run inside your existing controls. Read the threat-model doc →

How CyborgDB stacks up.

Capability

CyborgDB

Qdrant

Weaviate

pgvector

Milvus

Encrypted at rest

●

Encrypted during search

●

○

Filtered search

●

◐

●

Filters on encrypted metadata

●

○

Per-field / per-tenant key isolation

●

○

Bring-your-own KMS / HSM at index layer

●

○

◐

○

Key rotation without re-indexing

●

○

Crypto-shredding (delete = destroy key)

●

○

● full support · ◐ partial / conditional · ○ not supported

Encryption, without the tax.

Security doesn't have to mean compromising performance. CyborgDB keeps pace with unencrypted vector databases — and beats most of them.

Recall (%) vs queries per second across vector databases.
Database	Recall (%)	QPS
CyborgDB (encrypted)	72.8	792
CyborgDB (encrypted)	76.3	744
CyborgDB (encrypted)	81.3	672
CyborgDB (encrypted)	85.9	638
CyborgDB (encrypted)	88.8	567
CyborgDB (encrypted)	91.8	499
CyborgDB (encrypted)	94.0	433
CyborgDB (encrypted)	96.2	351
CyborgDB (encrypted)	97.3	321
CyborgDB (encrypted)	98.2	266
CyborgDB (encrypted)	98.6	245
CyborgDB (encrypted)	99.0	220
CyborgDB (encrypted)	99.4	182
CyborgDB (encrypted)	99.8	151
Qdrant	89.5	113
Qdrant	95.2	89
Qdrant	97.8	68
Qdrant	99.1	45
Qdrant	99.4	35
Qdrant	99.7	27
Qdrant	99.9	17
Weaviate	77.2	840
Weaviate	83.4	737
Weaviate	86.9	715
Weaviate	91.3	600
Weaviate	93.3	533
Weaviate	95.9	430
Weaviate	97.5	349
Weaviate	98.6	266
Weaviate	98.9	224
Weaviate	99.4	176
Weaviate	99.8	108
Milvus	92.6	66
Milvus	96.7	61
Milvus	98.3	55
Milvus	99.1	47
Milvus	99.5	41
Milvus	99.7	29
Elasticsearch	81.4	303
Elasticsearch	83.4	293
Elasticsearch	89.6	278
Elasticsearch	90.8	276
Elasticsearch	94.6	241
Elasticsearch	95.2	240
Elasticsearch	96.6	218
Elasticsearch	97.5	198
Elasticsearch	97.8	184
Elasticsearch	98.7	161
Elasticsearch	99.2	131
Elasticsearch	99.2	129
Elasticsearch	99.4	120
Elasticsearch	99.4	117
pgvector	81.4	835
pgvector	89.2	497
pgvector	94.1	402
pgvector	96.7	283
pgvector	97.6	190
pgvector	98.5	130
pgvector	99.1	82
pgvector	99.5	47
pgvector	99.6	38
LanceDB	94.2	282
LanceDB	97.5	174
LanceDB	99.0	101
LanceDB	99.3	90

CyborgDB encrypted

Qdrant

Weaviate

Milvus

Elasticsearch

pgvector

LanceDB

QPS @ 99% recall

CyborgDB v0.17.0 DiskIVF encrypted 220qps

Weaviate v1.37.0 HNSW 214qps

Elasticsearch v9.3.4 HNSW 143qps

LanceDB v0.27.1 IVF-PQ 101qps

pgvector v0.8.2 HNSW 90qps

Milvus v2.6.15 HNSW 48qps

Qdrant v1.17.1 HNSW 47qps

higher = better

DATASET wiki-all-1M · 768 dims · 1M vectors · top-k = 10

Index build time @ 99% recall

CyborgDB v0.17.0 DiskIVF encrypted 2m 32s

LanceDB v0.27.1 IVF-PQ 6m 06s

Qdrant v1.17.1 HNSW 20m 23s

Milvus v2.6.15 HNSW 28m 13s

Weaviate v1.37.0 HNSW 33m 30s

pgvector v0.8.2 HNSW 40m 05s

Elasticsearch v9.3.4 HNSW 65m 01s

lower = better

Single-threaded runs on the ann-benchmarks harness, c8g.4xlarge · May 2026.

Works with your existing stack.

CyborgDB is a binary you run inside the infrastructure you already operate. Nothing to migrate, nothing new to learn — just a privacy guarantee on top of the tools you already have.

Runtime

A drop-in Docker image.

Single container, multi-arch (amd64 / arm64). Helm chart for Kubernetes, Compose file for everything else. No TEE, no special silicon — runs anywhere Linux runs.

$ docker pull cyborg/cyborgdb:latest

$ helm install cyborgdb cyborg/cyborgdb

──────────────────────────────

linux/amd64 · linux/arm64

EKS · GKE · AKS · OpenShift · bare metal

FIPS 140-3 build available

Storage

Persists where you already do.

CyborgDB separates compute from storage. Point it at Postgres, MySQL, S3, GCS, Azure Blob, or any S3-compatible store. Your existing backup and DR plan covers it for free.

Postgres

AWS RDS

AWS S3

GCS

Azure Blob

MinIO

Cloudflare R2

Wasabi

Orchestration

Speaks the frameworks your team already uses.

First-party adapters for the orchestration frameworks. Same retriever interface, same indexing pipeline — swap the vector store and you're done.

LangChain

LlamaIndex

Haystack

Keys

Bring your KMS. Or your HSM.

Envelope encryption with the major cloud KMS providers and any PKCS#11 HSM. CyborgDB never sees your master key — only the data key, only at use, only in memory.

AWS KMS CMK + IAM

GCP KMS CryptoKey + IAM

Azure Key Vault managed HSM

HashiCorp Vault transit engine

Migrate your vector database.
Encrypted.

Already on Pinecone, Weaviate, or Milvus? Bring your workload to a thirty-minute call. We'll show you how CyborgDB performs against it. Free for up to 1M vectors.

Request a demo Start free

The vector database that stays encrypted even while it's searching.