Security

What we read, what we keep, who can see it.

Plain answers, each one true of the product today. Where something is not done yet, we say so.

What you get

  • We read only what you connect

    The night check reads an app only after you connect it — through our GitHub app, with that connection's own permission, or by uploading its code yourself.

  • Reading changes nothing

    The check downloads your repository's latest code once and reads it in memory; nothing is written to a disk.

  • A fix is made on a copy, and waits for your yes

    When you press Fix it, your own AI makes the change on a separate branch. Nothing reaches your main branch until you say yes.

  • Your AI key stays yours

    You connect your own model key and pay your model directly. The key is encrypted (AES-256) before it is stored, kept in a table that no signed-in session can read — ours, yours or anyone else's — and never shown again once saved, not even to you. Only our server opens it: to check that it works, and to run your job.

  • Files named like keys are not shown to the AI

    Files whose names say they hold credentials, such as .env, are withheld from the model, and the job's log names each one. A key pasted into ordinary code is not caught by its file's name — that is what the night check looks for.

  • Each workspace sees only its own

    Your apps, findings and receipts are fenced by workspace inside the database itself, so they are not readable from another customer's account.

  • Every yes leaves a receipt

    Who asked, which model made the change, what was checked, what it cost and who said yes. Each entry carries the fingerprint of the one before it, so a changed entry shows, and our app can only add to the record. Anyone can check a receipt themselves.

    Check a receipt

Technical details

We read only what you connect

There is no box where anyone types the name of someone else's repository. The App can read your code and write branches, pull requests and check results. It cannot touch your repository's administration, secrets, actions, workflows or members.

Reading changes nothing

It never opens your live database: which tables your code lets strangers read is worked out from your migrations. To look up known flaws it sends only the names and versions of the parts you use to the public flaw database, osv.dev — never your code. The check makes no AI calls. It needs no AI key — add one only when you want a fix.

Your AI key stays yours

Your code goes to the provider whose key you connected, when a fix is made, and our own server reads it to run the check and the fix. Beyond that, our administrative assistant (Claude, by Anthropic) reads it only for a support case you opened or during an incident, after a permanent record — see “Two different AIs” below.

  • Two different AIs

    Your app's AI work — reading your code, writing a fix — runs only on the provider whose key you connected, with your own key. Separately, Atrytone itself is run by one person, who can read every record in the database, and who uses an AI coding assistant, Claude (Anthropic), for administrative work. By default it is given counts, statuses and health, never your code, findings, messages or files. It reads those only for a support case you opened, or during an incident, and only after a permanent record names your workspace, the reason and what will be read. Anthropic is a recipient on that basis. More in our privacy page.

  • Where it lives

    Your account, findings and receipts are in a database hosted by Supabase in the EU (Frankfurt, eu-central-1). A fix job copies your code onto our server in Vienna, Austria, works on that copy and keeps a snapshot of the files it read and wrote with the job, so a stopped job can carry on. The copy on our server is deleted when the job ends; a daily check removes any copy left behind once it is 14 days old. The AI that makes a fix runs at the provider whose key you connected, wherever that provider runs it.

  • What we do not claim

    We do not promise that your app is safe, and using Atrytone does not, on its own, satisfy any law. The night check reads your code, not your running app, and every reading names what it could not read.

  • Report a security issue

    Found a hole in Atrytone itself, not in an app it checks? Email ops@atrytone.ai with what you found and how to reproduce it. A person reads it; there is no dedicated security address yet. Please do not test against another customer's workspace, and give us a reasonable time to fix it before telling anyone else. This address is also named in /.well-known/security.txt, the standard place a security researcher's own tools look first.