Don't trust us. Verify us.

A privacy claim you cannot check is worth nothing, however sincerely it is said, because anyone can say the same words. So this page does not ask for your trust. It gives you the public registers, the copy-paste commands and the honest gaps, so you can check us yourself in about five minutes, against sources we do not control.

Who we are, checkable

"Can't anyone register a company and an ICO number?" Yes. A company costs about £50 to form and ICO registration is a fee, not an inspection. Neither is a badge of merit and we will not pretend otherwise. What they actually buy you is accountability: a legal name on a public officer record, a registered office where court papers can be served, a regulator you can complain to for free, and filing obligations that leave a permanent public trail. An anonymous website that mishandles your data has broken a promise; a registered UK company whose director's legal name is on the record has committed an offence someone can act on. Anyone can register a company, but almost nobody puts the operator's real legal name on the officer record and then invites you to check it, unless they mean every word, because that turns every claim into a liability that follows a person, not a brand. So treat the registers as the floor, not the proof: they tell you who is accountable if the rest of this page turns out to be false. The rest of this page is how you check that it is not.

A named UK company with a named director is one you can complain about to the ICO or sue in a UK court. That is a stronger privacy promise than any adjective on any homepage.

And to be plain about scale: Hush is one founder and the machines he owns, not a corporate "we". You should learn that here, from us, rather than feel it was hidden. The privacy argument does not rest on headcount; it rests on the architecture and the registers above, which is why we would rather you check those than count us.

Check the connection yourself

Run these from any terminal. No account needed, nothing to install except curl (and Node.js for the last one). Expected results are as we observed them on 2026-07-26. If you ever get something worse, email [email protected]: you will have caught us drifting, and we want to know.

1. Security headers

curl -sI https://hush-ai.uk

Expect: strict-transport-security: max-age=63072000; includeSubDomains; preload (two years, on the browser preload list), a strict content-security-policy, x-frame-options: DENY, x-content-type-options: nosniff, referrer-policy: strict-origin-when-cross-origin, and a permissions-policy locking geolocation, microphone and camera.

You will also see server: cloudflare. That is real and we will not pretend otherwise: our public edge runs through Cloudflare for DDoS protection, and Cloudflare and Stripe (payments) are our only two sub-processors. Neither is an AI company. There is no AI sub-processor at all: inference happens only on machines we own in the UK.

2. The TLS floor

curl -s --tls-max 1.1 https://hush-ai.uk
curl -sI --tlsv1.3 https://hush-ai.uk | head -1

Expect: the first command fails (the handshake is refused; TLS 1.0 and 1.1 are off). The second returns HTTP/2 200. TLS 1.2 is still accepted so older clients can connect. We could write "TLS 1.3 only" and sound better; it would be false, so we do not.

3. The post-quantum handshake

node -e 'require("tls").connect({host:"hush-ai.uk",port:443,
  servername:"hush-ai.uk",groups:"X25519MLKEM768",
  minVersion:"TLSv1.3"},function(){
  console.log("PQ ok",this.getCipher().name);this.end()})'

Expect: PQ ok TLS_AES_256_GCM_SHA384. A handshake offering only the hybrid post-quantum group X25519MLKEM768 completes, so a PQ-capable browser gets harvest-now-decrypt-later protection on the connection today. The honest caveat: post-quantum key exchange is offered, not enforced. Classical key exchange is still accepted so older clients can connect. Available, not mandatory.

4. An independent grade, from someone who is not us

Run Qualys SSL Labs against hush-ai.uk yourself. We deliberately do not print our grade here: get it from them, not from us.

What we keep, what we do not, and how you check

What never lands on our servers

The content of your conversations and documents. It is processed in memory on machines we own, the answer is returned, and the content is gone. Any chat history you choose to keep lives encrypted on your own device, so deleting it there really deletes it. Your content is never used to train models and there is no third-party AI API in the chain to pass it to.

What we do keep, in plain sight

Account records, request metadata (which account called which endpoint when, never the content) and the financial records UK law makes us hold. The exact categories and retention periods are in the retention schedule in our privacy policy. A service that kept literally nothing could not bill you or stop abuse; a company that claims that is rounding, and rounding is how privacy marketing goes wrong.

How you check it, instead of believing it

Export your own audit log in one click, in CSV or JSON: your record, held by you, of every request that touched our servers. On a free pilot, your information-governance lead ends the two weeks with a report built from your own log, verifiable against your own work, not a testimonial from someone you cannot phone. The proof of what we hold sits on your side of the connection, which is exactly where it should sit.

What powers Hush

Open-weight AI models, used under open-source licences (stated plainly in clause 11 of our terms), served by standard open-source runtimes on hardware we own in the UK. No OpenAI, no Google, no Anthropic, no API resale in the chain.

You do not have to take that on faith either: every API response includes a model field naming the model that actually served your request, including when a busy lane falls back to a smaller model. A reseller marking up someone else's API cannot safely show you that field. Log it, compare it, hold us to it.

Two honest framings while we are here. First, these are efficient open models sized for professional drafting, summarising and filing work; if you want the largest frontier model on earth, that is not us. If you want confidential work done well off Big Tech servers, it is. Second, we will publish a signed-off public list of each lane's exact weights once we can keep that list reliably current; until then the per-response model field is the live truth, and it is the stronger check anyway.

The limit we will not spin. While a request is being answered, the model reads your text in memory. That is what answering means. So we never claim "zero-knowledge" and never claim we "cannot see" your data; no AI service without confidential-computing enclaves can honestly claim that, whatever their homepage says. Closing that final window takes enclaves with remote attestation. That is on our roadmap and it is not in production today, so we do not claim it. The difference between those two sentences is the difference between marketing and proof, and it is why this page exists.

Certifications, without rounding up

AssuranceStatusCheck or note
Companies House registrationHeldCompany 17278687
ICO registrationHeldRegistration ZC126901
Cyber EssentialsIn progressApplication in preparation via IASME. The certificate number and date will be printed here the day it is issued.
NHS DSPTIn progressSelf-assessment under way, not yet published.
Independent penetration testNot yet publishedA summary will be published here when an independent test has been commissioned and completed.
ISO 27001 / SOC 2Not heldNot certified. We say so rather than imply otherwise.

The full dated table, including insurance and the documents a DPO will ask for, is on the procurement page. Our Data Processing Agreement and Data Protection Impact Assessment exist as drafts with an accurate data flow; a real DPA is signed with your DPO before any live personal data flows, and we will not tell you otherwise.

Found a hole?

If any command on this page returns something worse than shown, or any claim anywhere on this site does not match what you observe, tell us at [email protected] (see security.txt). We would rather hear it from you first, and we fix what is reported.

Keep checking: the Trust Centre gathers every claim in one place; Security covers the architecture; Honest answers takes the hard questions; Procurement has the paperwork. Or skip ahead: start the free pilot and leave with your own audit log either way.