Mirror Security Product

VectaX.

Encrypted Memory and Encrypted Inference for AI Pipelines

VectaX solves the fundamental conflict between AI functionality and data privacy. Standard encryption breaks AI. VectaX uses Fully Homomorphic Encryption and similarity-preserving techniques so your models can operate on ciphertext without ever seeing plaintext data.

FHE
Core encryption
RBAC
Access control
FPE
Metadata encryption
3.9+
Python required
v1.0
SDK version
Overview

What is VectaX?

VectaX is Mirror Security's encryption layer for AI pipelines. It addresses a problem that no standard encryption tool solves: how to protect AI data without breaking AI functionality.

Standard AES encryption secures data at rest, but it destroys the mathematical properties that make machine learning work. You cannot compute cosine similarity between two encrypted vectors. You cannot run a model on an encrypted input. VectaX uses three specialised encryption techniques that preserve the properties AI needs while keeping data confidential end-to-end.

🔐
Fully Homomorphic Encryption
Compute on encrypted data without decrypting it. The model operates on ciphertext. No plaintext ever leaves your control.
〽️
Similarity-Preserving Encryption
Geometric relationships between vectors survive encryption. Nearest-neighbour search returns the same results on ciphertext as plaintext.
🏷️
Format-Preserving Encryption
Encrypt metadata, IDs, dates, and classification labels while keeping them queryable. No schema changes required.
🛡️
Multi-Dimensional RBAC
Decryption rights are scoped by role, group, and department simultaneously. Partial matches are denied. No side-channel access.
🔍
Encrypted Vector Search
ChromaDB-native integration. Search through millions of encrypted vectors with no performance penalty on search accuracy.
🔄
Secure Agents Pipeline
Protect every step of your ML pipeline from data ingestion through model serving. Inputs, outputs, prompts, and context all encrypted.
Encrypted Memory

The Problem with Plaintext Vector Stores

Every embedding is a compressed representation of the data that produced it. A RAG pipeline storing patient note embeddings holds a recoverable form of that data.

Any database breach, insider threat, or infrastructure compromise exposes the contents. Nearest-neighbour queries can be used to recover information about the original documents from their embeddings alone. Standard encryption solves the breach risk but breaks similarity search entirely. You end up choosing between security and the core feature that makes RAG useful. VectaX removes that choice.

How Encrypted Memory Works

VectaX encrypts embeddings so that relative distances between vectors are maintained in the ciphertext. A cosine similarity search on encrypted vectors returns the same ranked results as the same search on their plaintext originals. The vector database holds only ciphertext at all times.

Encrypted Memory Pipeline
Raw Data
Document / text
Embed
OpenAI / local encoder
Encrypt
FHE + similarity preserved
Store
Vector DB (ciphertext)
Search
Encrypted query
Decrypt
RBAC-authorised only
The vector database never holds plaintext at any step. Search accuracy is identical to unencrypted search.

Role-Based Access Control

Encrypting vectors is necessary but not sufficient. Without access control, anyone with the decryption key reads everything. VectaX ties decryption rights to organisational roles through a multi-dimensional key system.

Access keys are scoped simultaneously by role, group, and department. A user who matches two of three required dimensions is denied. Partial matches grant no access. A data scientist with model-training permissions does not automatically get access to patient record embeddings.

What a Policy Looks Like
A policy defines three dimensions at once: the role (e.g. clinical_team), the group (e.g. oncology), and the department (e.g. patient_records). A user key is generated per-user and scoped to their actual dimensions.
Enforcement at Decryption
The encrypted vector carries the policy it was encrypted under. Decryption is only possible with a user key that was generated under a matching policy. The enforcement is cryptographic, not application-level. It cannot be bypassed by changing database permissions.

Metadata Encryption with FPE

Most vector stores attach metadata to each embedding: document source, date, patient ID, classification tier. This metadata often contains the most sensitive fields. Format-Preserving Encryption encrypts this data while keeping its format intact, so existing database queries continue to work. No schema changes are required.

FPE uses a context-bound tweak generated from the metadata itself, ensuring that the encryption is specific to each record's context and cannot be replayed across records.

Standard RAG vs VectaX Encrypted RAG

Aspect Standard RAG VectaX Encrypted RAG
Embeddings at restPlaintext in vector DBEncrypted; similarity preserved
MetadataPlaintext JSON fieldsFPE-encrypted; still queryable
Document accessAnyone with DB credentialsRBAC-scoped per role, group, department
Similarity searchStandard cosine on plaintextSame results on ciphertext
Breach exposureFull embedding contents readableCiphertext only; no plaintext
Audit trailOptional, often absentPolicy-enforced, built-in
Compliance postureManual controls requiredEncryption at architecture level
Encrypted Inference

The Problem with Standard Inference

Inference is where sensitive data travels most. The model endpoint sees inputs and outputs in full. In cloud deployments, so does the provider.

A clinical model receives patient notes. A credit model receives financial history. A fraud model receives transaction details. Under GDPR, HIPAA, and sector-specific data residency rules, sending this data to a cloud inference endpoint in plaintext is often not acceptable. Teams respond by stripping sensitive fields before sending data, which loses accuracy, or keeping models on-premise, which loses scalability. Encrypted inference is a third option.

How Encrypted Inference Works

VectaX encrypts inputs before they leave your network. The model receives ciphertext, applies its weights using FHE, and returns ciphertext. Decryption happens only inside the authorised caller's context, under the RBAC policy set at inference time. The cloud provider, the model host, and the logging infrastructure hold ciphertext only at every stage.

Encrypted Inference Pipeline
Your Network
Raw Input
Plaintext at origin
Encrypt
FHE at client
Result
Plaintext output
Decrypt
RBAC-authorised
encrypted
encrypted
Cloud / Model Host
Receive Ciphertext
Plaintext never seen
FHE Compute
Weights applied to cipher
Return Ciphertext
Output also encrypted
The cloud provider operates on ciphertext at every stage. No decryption key is ever shared with the model host.

Pipeline Stage Comparison

StageStandard PipelineEncrypted Inference
Input preparationRaw data sent in plaintextEncrypted at client before any transmission
Feature extractionEmbeddings computed on plaintextComputed on encrypted input via FHE
Model computeWeights applied to plaintext featuresWeights applied to ciphertext; no decryption at host
OutputPlaintext prediction returnedEncrypted prediction; RBAC-gated decryption
LoggingRaw inputs and outputs in logsCiphertext only in logs; no plaintext exposure

Access Control Matrix

Role
Can Encrypt Input
Can Decrypt Output
Sees in Logs
Clinical team (dept-scoped)
Yes
Dept scope
Ciphertext
Data engineer
Test data
No
Ciphertext
Cloud / model host
Receives it
No
Ciphertext
Audit team
No
No
Cipher + metadata
Performance

Real Benchmarks: Plaintext vs FHE-Encrypted

Mirror's FHE engine is optimised specifically for AI workloads. These are production numbers, not theoretical estimates.

Performance
Plaintext vs FHE-Encrypted: Real Benchmarks
Single A100 / L40S-class GPU  |  FHE CKKS GEMV/Softmax
LLM Inference Throughput (tokens/sec)
Plaintext
Mirror FHE
7B Model150-240 tok/s
Plaintext: 200-300  |  FHE: 150-240
70B Model38-72 tok/s
Plaintext: 50-90  |  FHE: 38-72
MoE225-400 tok/s
Plaintext: 300-500  |  FHE: 225-400
TTFT — 7B (ms)50-150ms
Plaintext: 20-50ms  |  FHE: 50-150ms
Client Encrypt (ms)~2-5ms
Encryption overhead at client: ~2-5ms
AI Memory Quality (VectaX)
0.954
vs 0.97 plaintext
NDCG@5
98%
vs 99% plaintext
Temporal Acc
8ms
vs 4ms plaintext
Retrieval p95
4ms
vs 2ms plaintext
BM25 Search
Integrations
NVIDIA Nemo Memory NVIDIA Tensor RT NVIDIA Triton vLLM FasterTransformer All Vector DBs
Near-native speed with cryptographic guarantees. Memory quality loss under 2%. Inference overhead shrinking with each optimization cycle. Production-ready on A100/L40S-class GPUs.
Use Cases

Use Cases by Industry

Healthcare
Clinical RAG and Decision Support
Hospital RAG assistants store patient note embeddings encrypted. Physicians in cardiology cannot retrieve oncology records from the same collection. Clinical decision support models receive encrypted patient inputs; the cloud provider sees only ciphertext.
Financial Services
Trading Research and Credit Scoring
Trading firms store research note embeddings with RBAC by desk. Equities analysts cannot access fixed-income research. Credit scoring models operate on encrypted applicant data; the compute platform returns an encrypted score readable only by the credit team.
Legal
Matter-Scoped Document Search
Law firms embed client-confidential matter files scoped to the engagement team. Associates on Matter A cannot retrieve Matter B documents regardless of database permissions. FPE keeps matter IDs and dates queryable without exposing them.
Multi-tenant SaaS
Tenant-Isolated Embeddings
Each tenant's embeddings are encrypted under a tenant-specific policy. Tenants sharing the same vector collection cannot access each other's data even if the database infrastructure is shared or compromised.
Technical Reference

Technical Specifications

SDK packagespip install mirror-sdk mirror-enc
Python version3.9 and above
API endpointhttps://mirrorapi.azure-api.net/v1
AuthenticationMIRROR_API_KEY environment variable
Vector DB supportChromaDB (native integration); any store via standard interface
Encryption typesFHE (compute layer), similarity-preserving (vectors), FPE (metadata)
Access controlMulti-dimensional RBAC: role, group, department
Key scopingPer-user keys generated via sdk.rbac.generate_user_secret_key()
Metadata bindingContext-bound tweak via sdk.metadata.generate_tweak_from_data()
HostingAzure Static Web Apps (Academy); Mirror platform on Azure
SDK versionv1.0
Platform

How VectaX Fits with Other Mirror Products

VectaX
Data Encryption Layer
This product. Encrypts embeddings, metadata, prompts, and inference inputs and outputs end-to-end.
AgentIQ
Runtime Behaviour Layer
Monitors semantic content for injection, toxicity, PII, and hallucination. Operates on top of the encryption layer VectaX provides. The two products are designed to be deployed together.
DiscoveR
Security Testing Layer
Automated red teaming that finds weaknesses in the same AI applications VectaX protects. Use DiscoveR to validate that your VectaX-protected pipeline has no exploitable residual weaknesses.
Product Report
Download the VectaX Report
Get a complete PDF of this product page including architecture diagrams, capability tables, use cases, and technical specifications. Ready to share with your team or attach to a procurement brief.
PDF  |  A4  |  Mirror Security © 2025
academy.mirrorsecurity.io

Ready to encrypt your AI pipeline?

Request a personalised demo or explore the Mirror Academy to see VectaX in production-ready course examples.