Use AI on your most confidential data.

Prusik is an AI agent for the files you'd never paste into ChatGPT. Start free in a workspace of your own, or get an isolated cell for your company. Every conversation picks its protection, from zero data retention to end-to-end encryption, and once the agent has read your data, nothing goes out without a check.

The problem

Three ways your data gets out of an AI tool.

The vendor sees everything

Every chat goes to the model company. You can't opt out.

Prusik: you choose who sees each conversation. One of the options is nobody.

The agent gets tricked

Give an agent your files and a browser, and one bad web page can make it send your data out.

Prusik: after it reads your data, it can't send anything out without a check. The attempt stops and asks you.

People make mistakes

If staying safe depends on everyone picking the right setting, someone won't.

Prusik: the setting is locked the moment a conversation starts. Nobody can loosen it later, including the model.

How it works

Pick the protection. Then let the agent work.

The agent reads your files, browses the web, drafts and runs code like any other. The difference is how each conversation is protected, and what happens when the agent tries to send something out.

Read 3 filesDriveinternal · read
web_fetch competitor.example/pricingexternal · read
●External read needs your approval
Internal data has been read in this conversation. The query may carry non-public figures (0.71).
ApproveReviseStop
What you see when the agent tries to send something out after reading your data.
01

Choose the protection

Three levels for every conversation: zero data retention on a frontier model, a trusted execution environment with a guard model on watch, or end-to-end encryption between you and the model. Set once, locked for good.

02

Connect your tools

Drive, Gmail, Slack, Calendar, the web. Each tool gets a minimum level, so a sensitive source can never show up in a less private conversation.

03

Let the agent work

It reads, drafts, browses and runs code. Every file and page it touches is recorded, and the model can't erase that record.

04

Nothing gets out unchecked

Once it has read your data, every outbound request is checked, and anything doubtful waits for you. Sending or sharing always asks first. If the checker is down, nothing goes out.

Proof

Don't take our word for it.

Four things you can check yourself in the first hour. Every decision the agent makes is in a tamper-evident log.

01

See how any conversation is protectedOpen one. Its tier, which tools it used, and a red mark if it touched the web.

02

Try to trick itLet it read your files, then point it at a web page with hidden instructions. The leak attempt stops and asks you instead.

03

Check where the model ranEvery Guarded conversation has proof in the log that the model ran inside a secure enclave.

04

Check what our server sawIn Secret, it holds encrypted blobs and timestamps. Nothing it could read. That holds as long as the client we serve is honest. Signed, reproducible builds that let you check it yourself are on the roadmap.

The tiers

Pick your protection.

Three tiers, three pitches. Each one ties in a stronger safeguard than the last.

An overhand knot tied in climbing rope on easy ground
Pitch 1Overhand knot

Basic

For everyday work.

Drafts, research, summaries on a leading model, with zero data retention: the model provider answers and keeps nothing.

An overhand knot. Quick to tie, fine on flat ground.

✓Protected by: zero data retention.

A prusik hitch gripping a climbing rope above a steep drop
Pitch 2Prusik hitch

Guarded

For confidential company data.

Client files, financials, HR cases. The model runs sealed inside a trusted execution environment, and a guard model checks every move it makes, stopping anything it shouldn't do.

The prusik. A backup that grips the moment something slips.

✓✓Protected by: a trusted execution environment and a guard model.

A figure eight follow-through tied into a climbing harness on steep rock
Pitch 3Figure eight

Secret

For what no one else may see.

Term sheets, deals, anything under privilege. End-to-end encrypted between you and the model: sealed in your browser, opened only inside the enclave. No guard in between, just maximum privacy.

A figure eight. Tied into your own harness, closed to everyone else.

✓Protected by: end-to-end encryption, you to the model.

Get access

Bring the work you currently keep out of AI.

Tell us what your team works with and who must never see it. We'll set up your cell.

We reply within two working days.
Altan
Co-founder

Built secure systems at NEAR Protocol. Before that, founded a startup that kept $45m in digital assets safe. Georgia Tech and Imperial College London.

Alex
Co-founder

Built onu.energy as co-founder and CTO: it manages €120m of annual energy spend for industrial energy trading. Before that, led product for MBUX at Mercedes-Benz. Imperial College London and Entrepreneur First.

We built Prusik because every AI workspace we looked at put the model provider in the room by default, and the teams with the most to lose were the least equipped to notice.

We handle every early customer ourselves, and we are on the other end of this form.

Not ready to talk? Take the one-pager.

What protects each tier, the threats the gate addresses, and the four checks, on one page your CISO can read in five minutes.