Skip to content
Trausto

Das Whitepaper liegt nur auf Englisch vor. Zum Drucken oder Speichern als PDF: «Drucken» im Browser.

Trausto — Client-Side Encryption

Technical whitepaper for architecture, security and procurement review

Version 0.2 · 11 November 2026

Contact for questions on this document: security@trausto.com


What this document is, and what it is not

This is a description of how Trausto encrypts customer content, written for people whose job is to disbelieve it. It states the mechanism, the properties that follow from it, and — in the same detail — the properties that do not follow. Where the implementation is weaker than the architecture, the section says so rather than describing the target state.

It is not an audit. Nobody outside RiskZone GmbH has reviewed it at the time of writing. That review is planned as one of three independent evidence steps (external penetration test, external review of this document, ISO/IEC 27001 certification), and this page will name the reviewer and the date once it exists. Until then, treat this as a vendor statement that can be checked against the product, not as third-party assurance.

Derived from the internal architecture documentation (docs/security/e2ee.md), which remains the normative source. Where the two disagree, the internal document is correct and this one is stale — please report it.


1. The claim, in one paragraph

Assessment content — assets, zones, conduits, threat scenarios, risk ratings, uploaded evidence — is encrypted in the customer’s browser before it is transmitted. The server stores ciphertext and cannot decrypt it: it holds no private key, no data encryption key, and no recovery secret. A database compromise therefore yields ciphertext for the content we store. A rogue operator or a lawful order served on RiskZone GmbH yields the same for the content already stored; until a planned fix for project keys ships, such an operator can make a member use a project key the operator made, show that member content of its own as the project, and read what the member then writes to it. That does not extend to the application code our server delivers to the browser, which section 8 lists as not addressed. One exception exists and is described in section 6.

2. Key model

Element Mechanism
User identity key One active X25519 keypair per user, replaceable by an account-key rotation
Per-passkey protection Private key wrapped separately for each passkey via WebAuthn PRF
Content keys Per-project data encryption key (DEK), wrapped for each member
Recovery 24-word BIP39 phrase, Argon2id key derivation
Symmetric encryption AES-256-GCM
DEK wrapping AES-KW (256-bit) per member; AES-256-GCM under the organisation key
Key agreement derivation HKDF

One keypair, many devices. Each passkey has its own PRF-wrapped copy of the same private key, and each passkey is one entry in the user’s device list. A new device is approved from a device that is already signed in: after a fresh passkey confirmation there and a six-digit comparison of both screens, that browser encrypts the private key for a key only the new device holds. We relay that ciphertext and delete it once the device is set up. A passkey the platform syncs (iCloud Keychain, Google Password Manager) is one entry on every device it syncs to. Removing an entry revokes that copy — it is no longer served, but retained — on every device the passkey reaches. A lost device’s synced passkey keeps working on the user’s other devices until its entry is removed.

The unlock is bound to the passkey. The wrapping secret comes from a WebAuthn PRF ceremony against a passkey bound to the origin. The passkey’s secret never reaches us: a device-bound passkey keeps it on the device, and a synced passkey carries it in the platform provider’s end-to-end encrypted sync to the user’s other devices. It is phishing-resistant by construction: a credential registered for one origin does not respond to another.

Replacing the key after a suspected phrase leak. A new recovery phrase alone keeps the private key, so whoever copied the old phrase and its wrapped key would keep the key too. “My recovery phrase may have leaked” therefore replaces the key itself. Starting it needs a fresh passkey confirmation, never the phrase. The browser creates the new keypair and a new phrase and re-wraps every organisation and project key the user holds for the new key; we hold wrapped forms only. The change takes effect after 72 hours, announced to the account’s address with a cancel link; any signed-in session, the link or a recovery with the old phrase cancels it. Once the 72 hours have passed, it is applied in one of two ways. With the passkey that started it: that binds it to a passkey, not to a device — a passkey the platform syncs (iCloud Keychain, Google Password Manager) is the same passkey on every device it syncs to, and it can be applied from any of them. Or with the new recovery phrase, in a signed-in session on another device: the browser opens the new key with the phrase and checks it against the key the change set up, and the passkey that session signed in with — no other passkey, and only one that existed before the change was confirmed — must confirm the step. The phrase never leaves the browser. That confirmation moves the sign-in to a fresh session, so a copy of the old session cookie cannot take part in the apply. A failed attempt is recorded, and the first one is announced to the account’s address. A user who has neither the passkey that started the change nor the new phrase cannot apply it, and it expires. Applying it keeps only the passkey it was applied with, removes every other passkey of the account and signs out every other session; afterwards only the new phrase recovers the account. Every organisation and project key the user held is marked for renewal, and the next person who holds that key renews it when they open it. What the old key already opened stays readable to whoever kept it, as with any rotation (section 4). A change that is never applied expires after 30 days, and the old key stays.

Enterprise SSO is not offered today. When it comes, a federated login will prove identity only. It will not unwrap a key: the client will still perform the WebAuthn PRF ceremony or use the recovery phrase. Deriving key material from an ID token is excluded by design — an identity provider compromise must not become a content compromise.

3. Every ciphertext is bound to its purpose

All AES-GCM operations pass through one function that computes additional authenticated data as:

AAD = SHA-256( len(prefix)‖prefix ‖ len(purpose)‖purpose ‖ len(ownerId)‖ownerId
             ‖ len(bindingId)‖bindingId ‖ len(version)‖version )

Fields are length-prefixed with four big-endian bytes so that field boundaries cannot be shifted — ("ab","c") and ("a","bc") produce different AAD.

The practical consequence: a ciphertext produced for one purpose, owner or object fails to authenticate when presented in another position. A compromised server cannot serve object A’s ciphertext in place of object B and have the client accept it. A compromised server therefore cannot pass one object off as another. It can still serve an earlier state of the same object: the binding names the object, not its version. A project’s current data and each of its stored versions carry the same binding, so our server could present an older version as the current one, exchange one stored version for another, or withhold one, and the client would accept it. The same holds for the AI history, the risk register and the blueprint library.

Two constraints follow, and both are design rules we hold ourselves to:

4. Membership is a key operation, not a flag

Adding someone to a project wraps that project’s DEK for their key. Before the client wraps a key for another person, it checks the public key our server serves for them against the one it used before (section 8). When an administrator of the organisation removes someone who holds the key, the administrator’s client rotates the DEK and re-wraps it for the remaining members who may hold it: the project’s own members, and the organisation members the organisation key has been shared with. That rotation needs the administrator’s own copy of the project key; without one, the removal is refused. When the project’s owner removes them instead, nothing is rotated at once: the server marks the project as due for rotation, and an administrator renews the key. A member still waiting for that share gets no copy of the new key; once the organisation key is shared with them, they open the project through it. Deleting a project deletes its rows on our server, its key envelopes among them, and then its encrypted objects, the latter best effort. It destroys no key: the members’ own keys remain, a browser may still hold a cached envelope, and our database backups keep the deleted rows for their retention window. The content becomes unreadable because its encrypted objects are deleted, not because a key is.

The limit, stated plainly: what a removed party already decrypted and retained is outside the reach of any vendor. Rotation controls future access, not past copies. Anyone claiming otherwise is describing a system that cannot exist.

A reviewer reads everything and writes only review comments. A project can be shared with a reviewer — an auditor, for example — who holds the project key like any member: holding the key means reading everything, exports included. What keeps a reviewer from changing the analysis is our server, which accepts from them review comments and nothing else. Each comment and each later step on it (an edit, a resolution, a reopening, a withdrawal) is its own encrypted record, so the server can enforce the rules of a review trail: only a comment’s author edits or withdraws it, and edits it only while it is open; only editors and owners resolve or reopen it; no record is deleted. The one other write is a mark: whoever manages the project’s access can quarantine a record that no longer opens, which keeps it in the trail but out of the next key rotation. Authorship and the kind of each step are attested by us. The comment text is protected against us and against reviewers, but not against an editor or administrator who rotates the project key: a rotation re-encrypts every comment, and the person running it could change a comment’s text while its author and kind stay as we recorded them.

Removing an organisation member rotates the organisation-level key. Since v0.13.0-alpha (24 September 2026) the removal and the rotation are one operation. The removing administrator’s client generates the new organisation key and wraps it for every remaining holder, and the server accepts the removal only together with that payload. In one all-or-nothing batch, it then deletes the membership, replaces the member keys, re-wraps the project-level organisation wraps, re-encrypts the project titles, the risk register, the blueprint library and the organisation’s record of recipient key fingerprints, and revokes the departing member’s pending invitations to the organisation and those they sent to it. The server refuses a payload that hands the new key to the departing member (ROTATION_KEY_FOR_REMOVED_MEMBER). A removal whose rotation cannot commit leaves the membership untouched: it fails closed, never half-done. The organisation recovery envelope is re-wrapped in the same batch when an owner removes the member. After an administrator’s removal it is marked stale instead, and break-glass reads “needs regeneration” until an owner regenerates it. A removed member who cached the old organisation key holds a key that opens nothing the organisation writes after the removal.

What removal rotates below the organisation: the member’s key copies in every sub-organisation where they keep no role are deleted at once, and every project there whose key they held, and the sub-organisation’s own key if they held it, is marked on the server as due for rotation. The removing administrator often holds none of those keys, so the rotation is run by the next person who holds the key and opens the project (a sub-organisation’s own key by its administrators), with the same client-side rotation and a visible notice; only a completed rotation clears the mark. Until then the removed member has no access to the stored data, but what is saved meanwhile is still encrypted under a key they may have kept. The same mark covers a project of the organisation itself whose key the member held and that the removal could not rotate first: one the removing administrator holds no copy of, one whose data and AI history together are too large for one request, or one whose rotation failed; the notice after the removal names such a project. A member who deletes their own account, or whose account a tenant administrator resets, is not followed by a rotation either, but every project and organisation key the account held is marked the same way. There is no deprovisioning through an identity provider today. When SSO comes, such a deprovisioning will only be able to suspend a member, because completing a removal needs a key-holding administrator’s client to build the rotation. The manual rotation remains for a key exposed any other way.

5. Organisation-wide views are computed on the client

Risk aggregation across an organisation happens after decryption, in the browser, over exactly the projects the current user can decrypt. Two users with different access therefore see different organisation totals — which is correct, and occasionally surprising.

Server-side aggregation over decrypted content is out of scope by construction. Aggregate reporting to our server is not offered today. If it comes, it will be an explicit, organisation-administrator opt-in publishing a declared, minimal set of computed values — never the underlying content.

6. The exception: server-routed AI

Trausto offers an optional AI assistant. It has two working modes, and a third that can be configured but does not run today.

In-browser mode: the model runs client-side, in a browser that provides a built-in model, and only where a provider for this mode is set up. It needs the built-in model of a current desktop version of Chrome (its Prompt API). No content leaves the device for this purpose; we tested this in production on 2 November 2026, and the whole AI import ran on the device. A user who wants to guarantee that no content is ever server-routed selects in-browser mode; when the in-browser model fails, the request fails and nothing is sent to the server. The built-in model is small: treat what it proposes as a rough draft and review every entry before applying it.

Server-routed mode: the request — which may include file contents, entered text and current project data — passes through our worker in the clear to reach the one model the user selected and consented to. When it fails, no other model is tried: every request names its one model, the one the dialog shows, and only that model runs. The administration’s order of the models only decides which one the dialog preselects, and a model marked “explicit selection only” is never preselected. (Before app version 0.17, a request that named no model went through that order and tried the next model after a failure.) Each is either one of the platform models, reached through our AI gateway Opper (Sweden), which forwards the request to that model’s processor, or a model at the customer’s own endpoint. The tenant administration can disable any platform model. The gateway and the processors behind it are listed as our subprocessors on the website, with their locations and retention, next to the current list of platform models. For the duration of that request, that content is not encrypted from our perspective.

This is the one place where the sentence “the server never sees plaintext content” does not hold, and it is stated on the public website in the same words. The feature is opt-in. Operational telemetry about these calls (which provider and model, success or failure, duration, the token counts the provider reports, and for a failure an error code and our server’s error text) is retained; request content is not retained by us — what a model’s processor retains is stated with it in the subprocessor list.

If your threat model does not permit this, use in-browser mode or do not enable the assistant. We would rather lose the feature than the sentence.

Direct to the customer’s own endpoint: the tenant administration can also configure, through the API, a mode in which the browser calls the customer’s own AI endpoint directly — without our worker and without the consent dialog of the server-routed mode — and Trausto stores no credentials for that endpoint. Our content security policy blocks that call today, so such a request fails, and nothing is sent elsewhere instead.

7. What the server holds in the clear

Encryption is half a promise until you say what stays unencrypted — and for how long. Retention values are stated where they are fixed in code; a dash means the duration is operational and not asserted here. This table is the only copy: the website links here instead of repeating it.

Category Readable by us Retention Why it exists
Assessment content No — ciphertext only — Assets, zones, conduits, scenarios and evidence are encrypted on your device before upload. The one exception is the server-routed AI path below.
Uploaded evidence No — ciphertext only — Supplier documents and attachments follow the same path as assessment content.
AI processing Yes, for the duration of the request — only if you use the server-routed AI path — If you use the server-routed AI assistant, the content of that one request passes through our worker in the clear to the one model you select; when it fails, no other model is tried. It is either one of the platform models, reached through our AI gateway Opper (Sweden), or your own endpoint. Your tenant administration can disable any platform model; the processors behind the gateway are named, with location and retention, in our published subprocessor list. Running the model in the browser avoids this — that is the honest reason the option exists.
Key material No — wrapped only — Your private key is wrapped separately for each passkey, with a key derived in your browser from the passkey’s PRF output; your personal recovery phrase and the optional organisation one are generated in your browser and never leave it in the clear. We hold wrapped forms only. During a device approval we also hold the private key encrypted for the new device. We serve it for at most ten minutes, and delete it once the device is set up or, if it never is, in the nightly clean-up that follows. An organisation key can unwrap the projects in its permission chain — that is how an organisation keeps access when a person leaves, and it means the organisation recovery phrase reaches those projects too.
Recipient key fingerprints No — ciphertext only — One record per organisation, encrypted with the organisation key: for each person a key was wrapped for in that organisation, a fingerprint of their public key and its generation, so that every member’s device checks the keys our server serves against the ones the organisation used before (section 8). We see the record’s size, a version number, when it was created and last saved, and which member created it and which last saved it. Each save through the record’s own interface is in the audit log with its version and size; a rotation of the organisation key re-encrypts the record and logs that only as a yes/no flag, and a reset during such a rotation is logged on its own.
Risk register and blueprints No — ciphertext only — One record each per organisation, encrypted with the organisation key: the organisation’s risk register and its library of reusable analysis blueprints. We see each record’s size, a version number, when it was created and last saved, and which member created it and which last saved it. Each save through the record’s own interface is in the audit log with its version and size; a rotation of the organisation key re-encrypts each that exists and logs that only as a yes/no flag per record.
Account identity Yes — Your business email address and the interface language and region you chose. Needed to authenticate you and to address you in support.
Project membership Yes — structure, not content — Which record belongs to which project and organisation, who has access to it, and the names of your organisations and of your tenant. Project names and descriptions are encrypted with the organisation key; a project created before v0.7.6-alpha (19 July 2026) keeps its name and description in plaintext until it is migrated, which an organisation administrator’s opening does when the organisation key is unlocked and confirmed, as does a successful rename. That controls who can retrieve; the content becomes readable only through the project key.
Review comment records Yes — who, which comment, which step, when; never the text — Each review comment and each step on it is a record of its own, so the server can enforce who may write which step. We see per record: which project and comment it belongs to, who wrote it, whether it creates, edits, resolves, reopens or withdraws a comment, when, its identifier (for a comment converted from older project data, one that marks it as converted), which key generation encrypts it and its size, and whether and by whom it was quarantined. Each write is also in the audit log, with the writer’s project role, and so are reads of the trail. The text, the element it is about and a finding mark are encrypted.
Timestamps Yes — Created and last-changed times, so concurrent edits and version history work.
Audit log Yes 90 days (default) · SIEM export is the long-term archive Who did what, when. Required for the evidence trail the product exists to produce.
Usage volume Yes — Storage consumed and record counts, for billing and capacity.
Device and session metadata Yes Sessions valid 30 days, rows deleted in the first nightly sweep after they expire · device entries are revoked, not deleted Per device entry (one per passkey): the device name you chose (plain text), its public key, whether it is a synced passkey (the backup-eligible flag its authenticator reports), when it was added and when its wrapped key was last retrieved; removed devices are revoked, not deleted. Per session: browser identifier, an IP prefix — never the full address — and which device opened the session.
Transactional email Yes — sent via Mailgun (EU region) — Registration and recovery codes, invitations and account-change notices are delivered by the email provider Mailgun. The account-change notices tell a person that a change of their address was requested and when it takes effect, or that it was cancelled; that their address was changed; that a reset of their account was requested, or that it was cancelled; that their account was recovered or reset; that their account was deleted; that a passkey was added to or removed from it, or approved from another of their devices; that their recovery phrase was replaced; that a change of their account key was requested, cancelled, expired or applied; that their tenant role was changed; that an organisation or a project was transferred to them; that they were given access to a project; or that their access to an organisation was removed. These mails carry: confirmation codes, device name, organisation name (also in the subject line), role, the address of the acting person (a platform administrator is named only as “an administrator”), and the time of the change; on an address change, the old address is told the new one; the mail after an account recovery lists the passkeys the recovery retired, by device name and whether each was synced, or says that the other passkeys were kept. A mail about a project names its organisation, never the project, and carries no project content; where the project’s name is encrypted, the mail that gives access to a project says that it cannot show the name for that reason. Recipient addresses and mail metadata therefore also sit with the delivery provider.
Support and administration access Yes — metadata, never content — Platform administrators can read member lists including email addresses, audit logs and usage statistics; assessment content they cannot — it exists on our side only as ciphertext. Administrative writes are recorded in the audit log, and so are reads of member lists, audit data and platform statistics; other administrative reads (platform configuration) are not recorded individually. A platform administrator can also work in organisations and projects as an ordinary member, but only inside the shared platform tenant, never in a dedicated customer tenant; the administration console and every administrative write require a fresh WebAuthn step-up valid for at most 15 minutes, and administrative actions are tagged as such in the audit log. A tenant administrator can change a member’s address, and a platform administrator can change the address of a tenant’s only tenant administrator; the administrator is expected to check the person’s identity first. The change waits 72 hours and is announced to the old address, which can cancel it.
AI operational telemetry Yes — metadata — only if you use the server-routed AI path 90 days (default), as part of the audit log Which provider and model was called, whether it failed, how long it took and how many tokens it used; for a failure, an error code and our server’s error text. No request content.
SIEM export Yes — audit metadata — only with a webhook you configure — If your tenant administration configures a SIEM webhook, audit events go to the endpoint you name. Assessment content is not part of that stream.

Retention periods apply to the shared platform database. Dedicated databases are offered only on request and receive their own retention operation.

Data location. The ciphertext, the operational metadata and the audit log are stored in the platform databases (D1) and object storage (R2), which were created in Cloudflare’s EU jurisdiction (eu); that is binding and cannot be changed afterwards. Project content — the project data, its version history, the AI run history and uploaded documents — is stored as encrypted objects in object storage (R2); the database holds a reference to each object (its storage key, size and SHA-256), next to the metadata in the table above, the wrapped keys and the encrypted project names. Rate-limit counters and audit-deduplication markers are held in three places: in Cloudflare KV, in namespaces also created with jurisdiction eu (that binds their durable storage; KV can cache values outside the EU); in Cloudflare’s Rate Limiting service, which offers no data-location setting; and, for daily and hourly limits, in the platform database (D1). In KV and in the Rate Limiting service, IP addresses and account identifiers appear only as keyed hashes. Sessions are rows in the platform database (D1). The workers’ operational logs (Cloudflare Workers Logs) are kept by Cloudflare for a few days; among other things they can hold account, organisation and project identifiers and e-mail addresses shortened to their first and last two characters. Cloudflare states no storage location for these logs, so none is claimed, and none is claimed for the request metrics in Cloudflare’s Analytics Engine either. Delivery and DDoS protection run over Cloudflare’s global network, which stores no content. The operating company, RiskZone GmbH, is Swiss; Cloudflare offers no Swiss storage region, so none is claimed.

The metadata layer is a relationship register: which organisation, which people, which projects, who may see what. It is not nothing, and a buyer should weigh it as what it is.

8. Threat model

Addressed by the design

Not addressed, and why

9. Post-quantum status — read this carefully

The DEK envelope format is versioned and every version is tagged pre-quantum. Three exist: x25519-pre-pq-v1 and -v2 are retired: the client refuses them before any key agreement, because they did not bind the resource into the key derivation; x25519-pre-pq-v3 is the only version the product writes and reads, and it binds the recipient, the ephemeral key and the resource into the key derivation. In all three, the post-quantum input to the derivation is 32 zero bytes: a placeholder, fixed as a compatibility contract so a later hybrid derivation can be introduced without silently reinterpreting existing envelopes.

This means Trausto offers no post-quantum protection today. A hybrid version combining X25519 with ML-KEM is designed and its identifier reserved as x25519-mlkem768-hybrid-v1, but the write and unwrap paths are not shipped. Pre-quantum unwrap logic rejects envelopes claiming a post-quantum algorithm, so a mismatch fails loudly instead of degrading silently. Existing envelopes will require a controlled re-wrap when the hybrid version ships. The parameter set is decided: ML-KEM-768 (FIPS 203, NIST category 3), the set the reserved identifier names.

We state this because the reserved identifier is visible in the codebase and an unexplained reservation invites exactly the wrong inference. Any Trausto material claiming post-quantum protection before that migration completes is wrong, and we would like to hear about it.

10. What you can check without asking us

11. Data portability and exit

Export runs in the browser against the customer’s own keys. JSON carries the project’s current data in full and can be imported again. The archive, a ZIP file, carries that JSON together with the uploaded documentation files, the project’s AI history and its stored earlier versions, all decrypted in the browser; importing it brings back the data and the documentation files, while the AI history and the earlier versions stay in the archive. CSV and XML carry its tables for spreadsheets and other tools. It does not depend on our cooperation beyond the service being reachable. If the last key on the customer side is lost, the organisation recovery key can take over — provided the organisation has set it up and not disabled it, and one of its own owners (not an owner of a parent organisation) can still sign in to run the recovery with the offline recovery phrase. Neither we nor the customer’s tenant administrators can run it, an owner’s account cannot be reset, and only an owner can make someone an owner. An organisation whose only owner is also its last key holder therefore cannot use it once that owner has lost both passkey and recovery phrase; a second member who holds the organisation key is the protection against that. It is deliberately opt-in, because it lifts exactly the finality the rest of this document relies on. A user who has lost both passkey and recovery phrase cannot recover the account through the product. Their tenant’s administrator, who is expected to check the person’s identity first, can reset it, 72 hours after a notice to the account’s address, during which the account holder can cancel; the person then registers again with a new, empty account, and what was encrypted only to the old account stays unreadable. This works only for an account that owns nothing: as said above, an owner’s account cannot be reset.

Both are product functions, not contractual promises.


Open items named in this document

For a reviewer’s convenience, the places where implementation trails architecture:

  1. Post-quantum hybrid envelope (section 9) — designed, parameter set decided, identifier reserved, not shipped
  2. External review of this document (front matter) — planned, not yet commissioned

We publish this list rather than waiting until it is empty, because a whitepaper without one is either finished software or an incomplete document, and we are not finished software.


© 2026 RiskZone GmbH. Trausto is the product of RiskZone GmbH. Comments, corrections and disagreement: security@trausto.com