The answers, including
the awkward ones.
Written to be read by someone who is deciding whether to trust us with the worst thing in their week. Where the honest answer is “no” or “we cannot”, that is what it says.
The obvious ones
Can you read my conversations?
No — and not in the sense that we have a policy against it. Your vault key is derived from your passphrase with Argon2id inside your own session, and it only wraps a separate key for each conversation, document and project. Prompts leave your machine already encrypted to a key published inside the inference enclave's attestation quote, which your app verifies before it sends anything. There is no point in the path where we hold a key that opens your data.
Do you train on my data?
There is nothing to train on. We hold ciphertext we cannot open, and inference runs inside an enclave that does not retain what it sees. No corpus, no fine-tune, no “to improve our services” clause doing quiet work in the background.
What happens if I forget my passphrase?
You lose everything, permanently, and we cannot help you. That is not a gap in the product — it is the product. A recovery path is by definition a way in that does not need your passphrase, and anything that works for you works for anyone who can compel us. Write it down and keep the paper somewhere sensible.
Do I need an account?
No email, no phone number, no username. Your identity is a keypair generated on your machine. If you buy a plan our payment processor will have an email address to send a receipt to, and that address is never linked to a vault, a conversation or a key — it cannot be, because we do not know which vault is yours.
What if burnbox shuts down?
Your vault is on your own disk, encrypted with your own key, and the app opens it offline. Sync and the relay are conveniences, not dependencies. The relay itself is one docker compose command you can run yourself, and its build reproduces from source byte for byte so you can check the image is the code. The receipt verifier is a client-side signature check that works from a saved copy of the page.
What we can and cannot prove
Is it end-to-end encrypted all the way to the model?
No, and we will not say it is. Encryption terminates at the attested gateway, which forwards to a verified confidential host over a separately channel-bound hop. What we can show you: the enclave's measurement was verified — Intel TDX quote parsed, DCAP chain checked to the Intel root, report_data bound to a fresh nonce — before your prompt was encrypted to it, and the signed receipt records which host actually served the answer.
What happens if the enclave cannot be verified?
Nothing is sent. There is no send-anyway button, and a turn that cannot be encrypted is refused rather than quietly downgraded. This is the single most important sentence on this page: the failure mode is refusal, not leakage.
Does a burn really delete it?
Inside burnbox, yes. Destroying the data-encryption key makes the remaining ciphertext noise, and you get an Ed25519 receipt over the hash of what was destroyed. What it cannot reach are copies you made yourself: a screenshot, an exported file, a backup taken before the burn. The Privacy panel does best-effort shredding of local traces outside the vault and labels it best-effort, because on an SSD with copy-on-write and wear levelling that is the honest word.
Who can check a receipt?
Anyone holding it. The check is a canonical-JSON serialisation and an Ed25519 verification that runs entirely in the browser — no account, no app, no request to us. The desktop app and the web page import the same verification module, so the two can never disagree about what a valid receipt is.
What can't you prove?
The GPU. The gateway returns an empty attestation evidence list for every model we have probed while asserting a GPU architecture, so nobody can currently verify the GPU TEE claim — us included. We also cannot claim end-to-end encryption to the model, and the response-encryption spec does not cover tool-call arguments, so an agentic turn's tool arguments ride back unsealed. All of it is written up on the How it works page.
Money
Why is this more expensive than ChatGPT Plus?
Because you are not the other product. Confidential hardware is scarcer and slower to schedule than commodity GPUs, so an attested turn costs materially more to serve. And there is no second revenue line — no ads, no training data, no partnerships — so the subscription has to cover the whole thing. A cheaper subscription elsewhere is being subsidised by something; it is worth knowing by what.
Can I bring my own API key?
No, deliberately. Your own key means the model runs somewhere we cannot attest, with no enclave signature and no receipt — which removes the only reason to be here. It also re-introduces the identity anchor we spent the architecture removing, because a billing account is a real name. Bring-your-own keys are supported for data providers — markets, weather, news — since those are tool traffic, already itemised in the egress ledger, and never attested.
Do you offer SSO?
No, and it is not on the roadmap. Single sign-on makes your identity provider the thing that knows who uses burnbox and when. Half the point of this product is that no such thing exists anywhere.
Refunds?
Fourteen days, no questions, no exit interview. Cancel inside the app. Your vault stays readable offline forever — it is your disk and your key — and only attested inference stops.
Do you have my email address?
Only if you paid us, and then only because the payment processor needs somewhere to send an invoice. It is not linked to a vault, a conversation, a key or a device, and it is not used to identify you inside the product, because inside the product there is nothing to identify you against.
Practical
Which platforms?
macOS today. iOS is in active development and is a first-class target rather than an afterthought port — the Rust core and the React front end were built to move. Windows and Linux are honest maybes; we would rather ship two platforms properly than four badly.
Which models does it run?
Attested open-weight models inside Intel TDX enclaves, offered as Fast, Thinking and Auto tiers. The roster is chosen on what will actually attest and sign: we have removed models that turned out not to be what the roster claimed, and quarantined one that leaked. A menu entry here has to survive a signature check, not a marketing check.
Is it open source?
The parts that have to be verifiable are. The relay builds reproducibly from source and is meant to be self-hosted; the receipt verifier runs entirely in your browser where you can read it; the receipt format is documented so a third party can check one without our code. The desktop app is not public yet.
How do I get in?
Two doors. Buy a plan, or take an invite from someone already inside. Invites are unlimited and registered nowhere — enforcing a cap would mean the relay counting redemptions, which means the relay learning who invited whom, which is a social graph. So an invite is a capability, not a record: it cannot be counted and it cannot be revoked, because the mechanism that would allow either is the mechanism we refuse.
Can my whole newsroom or practice use it?
Yes — that is what the Scorched Earth tier is for: pooled allowance, group threads and shared drops, and your own relay, which we will either help you run or run dedicated for you. Talk to us at hello@burnbox.ai.
Ask us something harder.
Technical questions get technical answers, and if the answer is that we cannot do the thing, you will get that instead of a deflection.
