CIPHERBLOCK is the zero-knowledge privacy layer between your prompts and every LLM. Your instructions stay encrypted — end to end, in flight, and at rest.
Plaintext (local)
Summarize our Q4 revenue strategy
Ciphertext (wire)
▓▒░█▓▒░█▓▒░█
Trusted by 2,400+ AI-native teams
from seed to Series D
01 / The problem
Every system prompt, retrieval chain, and tool schema is a trade secret in plaintext. And every LLM provider logs, caches, or trains on something. The question isn't if your prompts leak — it's when.
73%
of enterprise AI features
leak prompts through third-party logs, fine-tuning datasets, or cached completions.
1 in 4
engineers admit
to pasting customer PII or proprietary code into an LLM without redaction.
$4.6M
avg. breach cost
for AI-adjacent incidents in 2025 — up 31% YoY per IBM Cost of a Breach.
At risk
System prompts
→ Logged by providers
At risk
User PII
→ Echoed in completions
At risk
Tool schemas
→ Scraped via jailbreaks
At risk
Retrieval context
→ Cached in CDNs
02 / How it works
STEP / 01
Drop-in SDK
Install cipherblock in 30 seconds. We wrap your existing fetch to OpenAI, Anthropic, Mistral, or any OpenAI-compatible endpoint.
import { wrap } from "cipherblock"
const openai = wrap(new OpenAI())STEP / 02
Zero-knowledge
Sensitive tokens are encrypted with AES-256-GCM on your device using keys we never see. Public tokens pass through for model reasoning.
// auto-detected & encrypted:
// {customer_email}, {revenue}, {secret_sauce}STEP / 03
One wire format
Ciphertext routes to any LLM. Responses are decrypted locally so only your app sees plaintext. Works with streaming, tools, and RAG.
const res = await openai.chat.complete( // response decrypted client-side )
Promise • no change to model output
Same quality. Private by default.
03 / Toolkit
Built for the way modern teams ship AI — from prototype to production, with a security posture your CISO will actually sign.
Envelope encryption with per-workspace master keys in AWS KMS, GCP KMS, or your own HSM. We operate the ciphertext; you operate the trust.
key: workspace.master
alg: AES-256-GCM
kms: aws/kms:us-east-1
$ cipherblock rotate --workspace=prod
Emails, card numbers, SSNs, IP, and free-form secrets get flagged and tokenized before they ever touch a model.
One SDK. 40+ providers. Switch with an env var.
allow / deny / redact by role, region, model, or cost.
Signed, append-only audit logs streamed to your SIEM. Stripe-quality observability for your prompt layer.
99.99%
uptime SLA
8.2ms
p50 overhead
7 yr
log retention
Hosted SaaS with SOC 2 Type II, single-tenant VPC, or fully on-prem Helm chart for regulated workloads.
Streaming-safe
Token-level encryption for SSE
Tool-call aware
Preserves JSON schemas
Semantic cache
Private hit/miss matching
Cost guardrails
Per-prompt spend limits
04 / Playground
Paste any prompt. Watch auto-redaction mark PII, then see the wire-format ciphertext your LLM actually receives. Nothing leaves your browser in this demo.
Plaintext / local
Auto-redact · detected 3
Refund customer ⟨EMAIL⟩ on card ⟨CARD⟩for order #A-2219. Notify her at ⟨PHONE⟩.
Ciphertext / wire
What your LLM provider sees
∷▓▫⟩⟩∎ ◆⟨◉◈▫⟪⟩⟫ ░▓░▓◈⟪⟪∷◉░⟨∷∎◯█◆∎ ▪⟪ ◉█∎░ ◆▫◯◆ ▓▪⟫█ ◎◈⟪░ ◉◇⟩▒ ⟫⟩◆ ◯▪▓▫∷ ◎▓◇◎◈⟩∎◇ ◯▒⟪∎⟪█ ▓◆◈ █⟨ ◯▒⟩⟨▫◯◆█⟪█⟩░█
AES-256
cipher
GCM
mode
256b
key len
Demo uses a visual substitution cipher. Production uses AES-256-GCM with envelope keys in your KMS.
05 / Pricing
Start free forever. Scale when you ship. No seat taxes, no surprise overages, no sales call to upgrade.
Hobby
For building & testing
Pro
For shipping teams
billed annually · $288/yr
No card · cancel anytime
Enterprise
For regulated workloads
30-day money-back guarantee. If CIPHERBLOCK doesn't reduce your prompt-leak surface by 10×, we'll refund you.
06 / Receipts
4.9
G2 rating
127
reviews
We had three months of security review before our GPT feature could ship. CIPHERBLOCK took that to three days. Legal signed on a Friday.
Maya Okafor
Head of AI, Northwind
Zero-knowledge was table stakes for our hospital customers. CIPHERBLOCK is the only layer we found that doesn’t ask us to trust a vendor with plaintext.
Jordan Reyes
Staff Eng · HeliosHealth
Drop-in SDK, sub-10ms overhead, and the audit log is prettier than our Stripe dashboard. This is how infra should feel.
Samir Patel
CTO · Layered AI
We switched from a homegrown proxy to CIPHERBLOCK in an afternoon. Reduced our prompt-leak surface to zero without changing a single app-layer call.
Elena Vogt
Principal Architect · Kosmos
07 / Questions
Yes. Master keys live in your KMS (AWS/GCP/Azure) or on-prem HSM. Our services only ever see ciphertext. We publish open-source attestation tooling so you can verify this claim on every release.
Median overhead is 8.2ms p50, 22ms p99. For streaming responses, encryption happens per token so users see no perceptible lag. We run global edge points in 19 regions.
OpenAI, Anthropic, Google (Gemini & Vertex), Mistral, Cohere, Groq, Together, Fireworks, AWS Bedrock, and anything with an OpenAI-compatible endpoint. SDKs for TypeScript, Python, Go, and Rust. Plugins for LangChain, LlamaIndex, and the Vercel AI SDK.
Yes. The Enterprise tier ships as a Helm chart with no external network dependency. Many of our regulated customers (health, finance, public sector) run fully air-gapped.
Nothing. Only sensitive entities are tokenized — the model reasons over placeholder tokens that preserve grammatical role. In our benchmarks, task quality is statistically identical to unencrypted baselines.
Per encrypted token, billed monthly. Hobby is free forever up to 10K tokens. Pro gives you 10M tokens + all features. No seats, no per-user fees, no overage traps — you get a hard spend cap.
Yes, on Enterprise. You own the keys, the logs, and the policy engine. We provide the same SDKs and a terraform module for air-gapped clouds.
Still curious? Chat with an engineer → book a 15-minute call.
08 / Ship it
Get early-access pricing, 10K tokens free every month, and a pair-programming session with our founding engineer. No card, no call required.
01
install SDK
02
wrap fetch
03
ship safe
Enterprise
Book a private demo
VPC · on-prem · HIPAA · 24/7 SE