Thorium | Keep confidential AI data inside

Confidential data is already reaching AI.

28.65m

secrets such as passwords and API keys leaked on public GitHub in 2025.

81%

rise in leaked secrets tied to AI services.

2×

the usual leak rate in public code co-written by an AI coding agent (3.2% against 1.5%).

Source: GitGuardian, State of Secrets Sprawl 2026 (March 2026). In March 2026 the NCSC warned that AI-assisted “vibe coding” is introducing new cyber-security risks.

Patterns miss business meaning.

Pattern tools are built for data that looks the same in every company. Your real secrets are ordinary sentences, and every organisation’s list is different.

What pattern tools catch

4539 1488 0343 6467QQ 12 34 56 Csk_live_…[email protected]

Card numbers, National Insurance numbers, keys and addresses: fixed shapes a regular expression can find.

What they miss, and Thorium learns

Harbourline is moving its £4m book to us in Q1 if we hold 35 basis points.client terms

The Atlas launch moves to the first Tuesday after the regulator call.unreleased product

The board expects the quarter to land six million below the external range.board figures

Please schedule a regular team meeting for Friday afternoon.safe

Illustrative examples written for testing.

Every AI request checked locally.

Three checks, one decisionFixed patterns, a small local language model that recognises confidential meaning, and your own categories. The strongest signal decides.

A learning loopEvery block is logged. An administrator reviews it and confirms or corrects it, and your categories update. The example text is discarded.

Allowed requests onlyClean requests go to OpenAI, Anthropic or Google as normal. Blocked requests stop at your boundary and the caller is told why.

People and agents alikeCoding agents and apps use the learned gateway and shared dashboard. The browser extension currently adds local pattern protection on supported AI websites.

Teach Thorium what is confidential.

Detect

A request is checked

Thorium blocks or warns, and logs the event as a hash.

Review

An administrator reviews it

They confirm the decision, or mark it as safe or confidential.

Improve

The next decision is better

Six corrections took accuracy on new wording from 50% to 87.5% in testing.

Update

Your categories update

Versioned and isolated to your organisation. No retraining run needed.

Teach a new category · example

confidential“Project Northstar ships to the pilot bank in March.”

confidential“Northstar pricing stays under wraps until launch.”

safe“The north office is closed on Friday.”

safe“Our public roadmap is on the website.”

◆ Category active from 2 + 2 examples What Thorium keeps

SHA-256 hashnumeric embeddingexample text

Why was this prompt blocked?

A policy says what should happen. Thorium’s shared dashboard shows what did, without storing the original confidential text.

thorium · incident logexample view

Shared dashboardGateway incidents appear in one place for review, policy decisions and follow-up.

Tenant-isolated learningEach organisation’s corrections update only its own versioned decision boundary.

Hash, never contentEach incident records who, which tool and why. The log cannot become a leak itself.

Your infrastructure. Your keys.

# on a server inside your network (example)
$ docker compose up -d
ready nginx       published · the only exposed service
ready dlp-engine  internal network
ready router      internal network
ready gateway     internal network
ready dashboard   internal network
ready langfuse    internal network
ready postgres    internal network

# point tools and agents at Thorium instead of the provider
$ export OPENAI_BASE_URL=https://thorium.yourfirm.internal/v1
$ export OPENAI_API_KEY=<your Thorium key>

Self-hosted as a Docker Compose stack. One published entry point; every other service sits on an internal network.
Standard hardware. The learned engine runs on CPU. No GPU, no external classification service.
Your keys, your models. Staff keep using current models from OpenAI, Anthropic and Google.
Per-agent control. Each agent gets its own key, model rules and spending limit.

Proven in repeatable tests.

104 / 104

confidential messages caught in the synthetic benchmark, with no false alarms on 100 safe messages. The illustrative test scanner caught 4; the deployed Thorium scanners caught 22 on the same set.

50% → 87.5%

accuracy on new, differently worded messages after six administrator corrections.

0.97 ms

median added by the learned engine per message, on a standard CPU. A model’s own reply takes seconds.

Frozen test setsFixed before each run and versioned.

Automated runsRe-run from scratch on a clean machine in GitHub Actions.

Unseen messagesTested on messages the model never saw while learning.

Your pilot, your dataAccuracy on your own data is measured in the pilot.

Controlled English lab test using non-customer examples, October 2026. Pilot accuracy is measured separately on each customer’s approved test set.

Published prices. No enterprise contract.

Core

£6

per user per month, from 25 users

Pattern and generic semantic detection
Every AI request checked, with an audit log

Start a 30-day trial

Business

£12

per user per month, from 50 users, plus £1,500 onboarding

Everything in Core
Learned detection of your own confidential categories
Shared incident dashboard and versioned policy evidence

Start a free pilot

Regulated

£20,000

a year and up, plus £5,000 onboarding

Everything in Business
Service-level agreement
Leakage-test report for each deployment

Start a free pilot

Annual billing. Prices exclude VAT.

Pilot Thorium in your environment.

Install

We help you install Thorium on your own infrastructure.

Teach

You give it a handful of examples of what is confidential to you.

Review

You compare what it catches, review the dashboard and decide what to enforce.

Built for UK organisations of 25 to 249 people that use AI tools or coding agents and need confidential prompts to stay inside their control.

Thorium is inviting its first pilot customers · [email protected]