Security
How Hesimesi protects what you store, what our server can and can't read, and the limits we know about. Hesimesi is in an invite-only beta: anyone can sign up, but an account opens only when we approve it by hand. This page is candid about what is not done yet.
Last updated 21 November 2026
In short
- What you save or lock is sealed on your device, and our server can't open it. The exception is a file you lock only to a date: once that date passes, anyone who holds it, us included, can open it. We can read what is around it: your email address, file names, labels, your sign-in log and the rest listed below.
- Because we can't open it, we can't recover it if you lose your passphrase. Recovery shares held by people you trust are the answer.
- The limits we know about: most approvals rest on your signed-in session, the cryptography can't protect you if we serve you malicious code, and no outside security audit has been done.
- Report a security problem to security@hesimesi.com. We acknowledge within 5 business days.
The design in plain words
When you save a detail or lock a file, it is sealed on your own device, in your browser, before anything is sent to us. Sealing encrypts it with a random key, and that key is itself locked in one or more ways you choose: with a passphrase (your vault passphrase for saved details, or one you pick for a file), with recovery shares, with an unlock date, or with the key of your sidecar, a small helper program you run on your own computer or server to deliver values. We store only the sealed copy.
Your vault passphrase is used in your browser and is not sent to us, as long as the code our website sends you is honest (see “Honest limits” below). Your account password is a separate thing: we keep only a one-way hash of it, not the password itself.
We don't implement any cryptographic algorithm ourselves. The engine combines established, published algorithms through existing implementations: XChaCha20-Poly1305 to encrypt, Argon2id to turn a passphrase into a key, HKDF-SHA-256 to derive keys, and Shamir secret sharing for recovery. A delivery to a site is sealed on your device to that site's own key with HPKE, using a hybrid of a classical algorithm (X25519) and a post-quantum one (ML-KEM-768).
Our server program is built without the sealing engine. A saved detail is opened only where it is meant to be used: in your browser or the approval app; in your sidecar, on your own computer or server; or, for a delivery to a site, on that site's own server, with its own key. A locked file can be opened by anyone who holds both the file and what unlocks it: its passphrase, enough recovery shares, or, for a file locked only to a date, that date having passed.
Because we can't open what you sealed, we also can't recover it if you lose your passphrase. Recovery is the answer: it splits a spare key into recovery shares you give to people you trust, your trustees. Enough of them together let you back in; fewer than that open nothing.
What we can read is listed in the next section.
What our server holds and can read
Our server stores sealed copies of your files and saved details, which it can't open, except a file locked only to a date once that date has passed. Everything else it holds is readable to us, including:
- your account email, and the hash of your account password;
- the names of your files (a file's original name), and the labels and field names of your saved details;
- the destination addresses you add for deliveries;
- the sign-in log: each recorded sign-in or attempt, with its full IP address and the account's email;
- your activity log, which records what you and your agents did;
- messages sent through our contact form, with the sender's IP address;
- the push address your browser's push service gives us when you turn on approval notifications.
What a breach of our storage would expose
A breach of our storage would expose the readable details above, but not the contents of what you sealed, except a file you locked only to a date once that date has passed, which anyone holding it can open. For a delivery to a site, the sealed bundle stays only in the server's memory, for at most about 21 minutes, and is not written to its database.
The privacy page lists everything we keep, why, who receives it and for how long.
Links we email or share
Every one-time link we send (sign-in links, email confirmations, passkey links, invites and share links) carries its secret after the # sign. Browsers don't send that part of a link to any server, so the secret is not in the page addresses our hosting providers' request logs record. When the page uses the secret, it sends it to us in the body of a request, not in an address. Our server's own log lines record only which kind of request it was, not its address or contents.
Our email sender keeps a copy of each email we send, links included, for up to 30 days, plus up to 7 days in its backups; the privacy page lists it. Most of these links work only once.
A mail service that rewrites links, or a browser that syncs your history to another service, still sees the whole link. Treat these links as private and don't forward them.
Signing in
- You sign in with your password, a passkey, or a sign-in link we email you.
- By default your browser usually drops the sign-in when it closes (one that restores your tabs may keep it), and our side ends it after at most 12 hours. Tick “Keep me signed in” to stay signed in for 30 days.
- Repeated wrong passwords for one email pause password sign-in for that email for a while, except in a browser that has signed in to that account with the right password before, which has an allowance of its own. Signing in with a passkey, or with a sign-in link on the website, still works.
- Adding or removing a passkey needs a fresh proof in the same session: an existing passkey, or your password followed by a link we email you.
- We email you about every new passkey, every removal and every password sign-in. After the fourth password sign-in in one hour, the rest of that hour gets no separate mail, and the fourth mail tells you so.
- In Settings you can see your recent sign-ins and sign out every other device at once.
The sign-in log
We record each sign-in and sign-up, each wrong password that gets past our rate limits, and each refused sign-in to a suspended account. A record holds the time, the method (password, sign-in link, passkey or sign-up), the outcome, the full IP address, and the account's email at that moment, with whether it had been confirmed. A wrong password for an email that has no account keeps the time, the method and the IP address, but no email.
We don't record attempts our rate limits refuse, failed passkey prompts, or invalid or expired sign-in links.
Recording is best effort. If our database fails at that moment, a record can be lost, and your sign-in is answered exactly as it would have been.
You can see your 50 most recent records in Settings, wrong passwords included. A wrong password recorded before your email was confirmed is shown without its IP address, because only the typed email tied it to your account.
Admins with the permission can see any account's records, to stop abuse and to answer lawful requests. Each such look is written to our record of admin actions before anything is shown.
Our server's own log lines carry no IP addresses and no emails; the database's error lines can hold an email address, as the privacy page explains. Inside our service, the sign-in log is the only place we keep the IP address of a sign-in; our hosting providers' request logs also include IP addresses, as the privacy page explains.
For now we keep sign-in records with no fixed end date, also after an account is deleted. The privacy page says why, and that a fixed period will be set before public launch, which means when sign-up opens without our approval, the first customer pays, or a company is registered, whichever comes first.
Suspending an account
An admin can suspend an account to stop abuse. A person decides, and records a reason that is kept with the suspension. At that moment every session, agent key, sidecar and share link of the account is cut off, and each is refused if used later, as are the API keys of organisations the account owns. Deliveries no sidecar has picked up yet are cancelled, a site's open requests from the account's agents end within about a minute, counted as denied, and agent requests waiting for approval can't be approved while the suspension lasts. Each of those expires 24 to 30 hours after it was made; if the suspension is lifted first, the owner can still approve it. A site registered under one of the account's organisations stops working too, including deliveries already approved for other people.
A suspended account can't sign in: after a correct password, sign-in link or passkey it sees “This account is suspended. Contact support@hesimesi.com.” Anyone who doesn't hold the credential gets the usual failure answer, so a suspension is not revealed to them.
A suspension deletes none of your files, saved details or account data; deliveries not yet picked up are cancelled. Lifting it restores nothing that was cut off: the owner signs in again and creates new keys and links, while the organisations' API keys work again. Every suspension and every lift is written to our record of admin actions.
Honest limits
- Most approvals rest on your signed-in session. Approving a delivery to a site needs your passkey, but approving a send, a release to a destination or a browser fill needs only the signed-in session, so an AI agent that controls your signed-in browser could approve its own request. (A high-assurance release also needs your passphrase or enough recovery shares.) We plan to require a passkey for these approvals before sign-up opens without our approval.
- Someone with write access to our database could alter the activity log. Its entries are chained, which catches casual changes, but on this deployment the chain is not protected by a secret key, so such a person could rewrite entries without it showing, or cut entries off its end.
- The cryptography doesn't protect you against us serving you malicious code. The code that seals your data is sent by our website each time you load it, so a compromised or coerced deployment could send code that captures your passphrase. There is no independent check against this yet. What the cryptography does protect you against is a breach of our storage, and us reading what we store, except a file you locked only to a date once that date has passed.
- A compromised browser on your device can read what you seal there.
- A guessable passphrase can be guessed; no setting makes up for it.
- Your sidecar takes which value to deliver, and when, from our server. Someone who controlled our server or its database could make a running sidecar release a value saved at the standard level (the level your sidecar can open on its own), to a destination already on that sidecar's own list, with no request and no approval. The activity log would show it only afterwards.
- A value saved at the high-assurance level opens only with your passphrase or enough recovery shares. If you lose your passphrase and more of your trustees than your recovery allows, nobody can recover it, including us.
- A verified destination has proved it controls its address, not that it handles your value well.
- Our public pages, the vault, the approval app and the admin console are served from one web address, so a script injected into any of them could act with your session there.
- Suspension relies on sign-up needing our approval: a suspended person could sign up again with another email, and only that approval step stops them.
- No outside security audit has been done. Hesimesi has had internal reviews only, which is a weaker thing. The post-quantum library behind site deliveries has not had an outside audit either.
- One person maintains Hesimesi, and there is no registered company behind it yet.
Our threat model, in short
Our threat model is written for someone deciding whether to trust Hesimesi with real data. Its central claim is narrow: our server can't open what you sealed, and the few places where a readable copy does exist are listed. It doesn't claim that nobody ever sees a value. It covers saved details that AI agents use, and deliveries to sites. It doesn't yet cover time-locked files, locking a single file, or the business composition panel, which each work differently.
How far an AI agent is kept from a value depends on how the value is delivered. In a delivery to a site, the agent carries only a single-use token, not the values. That holds on three conditions: you approve on a device or browser profile the agent doesn't drive, the site never puts a delivered value back into the page, and our website's code reaches you unmodified. When the browser extension fills a value into a page, no answer carries the value to the agent, but an agent that drives that browser can read the page; a fill reduces what the agent sees, and doesn't guarantee it.
Where a readable copy exists:
- your browser, while you type and seal a value; anything that controls that browser, an AI agent included, can catch the value and your passphrase as you type them, so seal values in a browser no agent drives;
- your sidecar, on your own computer or server, while it delivers a value, and the destination that receives it;
- your approval app, while you pick or type values for a delivery to a site, and that site's own server once it opens the delivery;
- the readable details listed under “What our server holds and can read”.
Reporting a security problem
Email security@hesimesi.com. Our source code isn't public yet, so email is the only channel. For bugs and questions that aren't about security, write to support@hesimesi.com.
Please don't publish a finding before the disclosure window below has run. Most useful to include, roughly in this order:
- what an attacker gains;
- the smallest reproduction you have;
- where you saw it: the page or API route, and its host;
- when: the date and time, with the time zone;
- what you used: the browser and its version, or which of our apps or tools and its version.
What happens after you report
Acknowledgement
- Our commitment
- Within 5 business days
Initial assessment, with a severity
- Our commitment
- Within 10 business days
Fix for a problem accepted as High or Critical
- Our commitment
- Target: 90 days
Coordinated disclosure
- Our commitment
- 90 days from the acknowledgement
If we are late
The 90-day window binds us, not you. If the fix isn't out by then, publish anyway; if it ships sooner, publish as soon as it ships. If we haven't acknowledged your report within 5 business days, treat that as our failure and disclose as you see fit.
What you may test
- The hosted beta at www.hesimesi.com (with the approval app at /approve/) and api.hesimesi.com, used the way an approved tester uses it, with your own account: the vault, the browser extension, the approval app, an agent you connect, the downloadable tools, and our demo visa site with sample details.
- Beyond that use, inspection only: the responses, headers, content security policy, public pages and shipped scripts of those hosts, and the demo visa site with its public declaration file. Approved testers can ask support@hesimesi.com for the demo site's address.
- Not allowed: automated scanning, fuzzing, brute force, load or denial-of-service testing, or any attempt to reach another user's data. Use sample details only.
- Out of scope: any other host; the providers our service runs on (the server host, which also runs our database, the web host, file storage, the email sender, the inbound email service, the DNS provider and the domain registrar), which you report to directly; physical attacks, social engineering and anything aimed at a person; and flaws in third-party code, which belong upstream, unless the flaw is in how we use it.
Safe harbour
Research done in line with this policy is authorised. We won't treat it as a breach of our terms, and we won't take legal action over it.
Two honest limits. The promise comes from the individual who runs Hesimesi, not a company, so it binds them but can't bind the server host, the web host, file storage, the email sender, the inbound email service or the DNS provider, whose own terms apply to anything that touches their systems. And it covers only what this policy allows: nothing against another person's data or another party's systems.
Act in good faith, stay in scope, avoid privacy violations and service disruption, and use only accounts you own.
No bug bounty
We can't offer payment: there is no company yet. What we offer instead: credit in our public thanks list once the source is published (or anonymity, if you prefer), a written reference describing your work, and help requesting a public vulnerability identifier where one is warranted. Until the source is published, the written reference is the credit you can show.
Changes to this page
Before a change that affects your data or your rights, we email every account. The date at the top shows the latest edit.