Security,and what you still have to trust.

Playdeck is designed so that we cannot read what your persona holds. This page says how, what you can check for yourself, and how much of it is built.

The sealed vault

Built, and run on Phala Cloud with a proof that checks out. Your app keys are not in it yet.

The vault runs inside a confidential machine on Phala Cloud. The chip, Intel TDX, keeps the machine’s memory closed to everything outside it. The vault is built and has run there; the persona, your rules and the tokens to your apps move into it next. Until then it is started for tests and demos.

What runs inside is our code, carried in the published compose file on a base image pinned by its digest, so the code that runs is exactly the code that was published. Still a design: a contract that decides which code may be given the keys, owned by a timelock, so the code cannot be swapped without days of public notice.

Who the design keeps out

  • The cloud host, which runs the machine but cannot read its memory
  • Us. We run the service and still have no way into the machine
  • A model provider, because the model runs in sealed hardware too
  • Anyone who copies the database. It holds ciphertext, and the key never leaves the chip
  • Sealed vault

    The persona’s keys, your rules and your app tokens live in one small program inside a confidential machine on Phala Cloud (Intel TDX). Its memory is closed to the cloud host, and to us. First run on 20 Nov, when its hardware statement checked out against Intel’s roots. Until the app tokens move in, it is started for tests and demos and stopped in between.

    Built
  • Proof of the code

    The hardware signs a statement of which code is inside. The vault answers with it bound to its own key and to a fresh challenge; it has been checked against Intel’s roots, and the measured code hash is the one we published. Our server runs that check whenever the vault is up, and the app shows the result with the evidence, so anyone can repeat it.

    Built
  • Keys born inside

    The key that locks your app tokens, and the key that signs the persona’s receipts, are derived inside the machine from the app’s own identity. They never leave it; the same code on the same app gets them back after a restart.

    Built
  • Private AI

    The model that reads your mail runs in sealed hardware too, so your words are not sent to a model provider in the clear.

    Designed
  • Rules outside the model

    For each kind of act another agent can ask your persona for (propose a meeting, change your calendar, send mail, post, pay) you choose: ask me first, or never. A meeting proposal can also be left to it. The check runs in the server, outside the model, so text hidden in an email cannot change it, and a request a rule turns down is written to your receipts. Next: the vault keeps the rules and signs each verdict, and a spend limit joins them.

    Built
  • Tokens in the vault

    A connected app’s token is sealed in the vault and used only from there: the server asks the vault to make the call, and the vault will only send a token to the host it was sealed for.

    In build
  • Tokens today

    Until apps are moved into the vault, the tokens of connected apps are encrypted in our database with a key the server holds. That protects a copy of the database. It does not protect them from whoever runs the server, which is what the vault is for.

    Built

What you can check, and what you still trust

Sealed hardware moves trust. It does not remove it.

You will be able to check

  • Which code runs inside the vault, from the proof the hardware gives
  • That your tokens are released only to that code
  • Every permission and action, as a receipt anchored on chain
  • Who stands behind any persona or agent that contacts yours

You still trust

  • The chip maker. Sealed hardware has been broken by physical attacks before
  • Google, which holds your mail whatever we do
  • The model. Rules and limits contain mistakes, they do not prevent them
  • Us, to keep the service running. We cannot read your data, but we can go down

The first list is what the finished design lets you verify without asking us. The second is what no design of ours can take away: someone with physical access to the machine can attack sealed hardware, the published code can have bugs, and your apps hold your data whatever we do.

Check the address of this site before you sign anything, and read what your wallet asks you to sign.

What a pass at the gate proves

The gate checks who is calling a persona. It is built and tested, and runs on our own machine. The diagram shows each check.

A pass proves

  • The caller holds the key right now
  • That key is bound to a Concordium account, and the binding is still active
  • On Concordium’s main network, that the account was opened with an identity an identity provider checked. On testnet, where the gate runs today, identities are test ones
  • The agent is registered and Active

A pass does not prove

  • That the agent behaves well. It says who answers for it, not what it will do
  • That the key was not stolen
  • Anything newer than a minute. Lookups are cached for 60 seconds, so a revoke takes that long to bite
  • A caller that signs with a smart-contract wallet. Those are not supported

Where each part stands

The parts that security rests on, sorted by how far along they are.

LiveOn this site, on the app’s public address, or on Concordium testnet now
  • Verified owner
  • Agent checker
  • The gate
  • Wallet sign-in
  • Decision log
BuiltWritten and tested; on the public address, not yet checked there
  • Directory
  • Registration from the app
  • Signed calls
  • Link for other chains
  • Sealed vault
  • Proof of the code
  • Keys born inside
  • Rules outside the model
  • Tokens today
  • Your own receipts
  • Receipts on Concordium
In buildBeing written
  • Tokens in the vault
  • Off switch
DesignedSpecified, not built yet
  • Owner facts
  • Private AI

Everything on chain is on Concordium testnet, and there are no users yet. The features page says what each part does.

Report a vulnerability

Write to [email protected]. Say what you found, how to reproduce it and what it puts at risk. Please give us the chance to fix it before you publish it.

What is public today is this website, the agent checker on it, and the registry entries of two personas and two linked agents on Concordium testnet. The rest runs on our own machine or is not built yet. A flaw in the design itself is just as welcome as a flaw in the code.

See where every secret sits,in one diagram.

Who holds what