Security overview

The short version

What every customer gets:

• Every customer gets their own model. It learns from that customer's work, never another customer's.
• Every customer's code and model live in storage of their own, under an encryption key of their own.
• We don't train Airgetlam's own models on any customer's data, on any plan. Your contract says so, and only a change your company signs could alter it.
• You choose which repositories and paths Ryo may learn from. Every settings change is logged with who made it and when, and so is each addition or removal of repositories — how many, never which.

What we are still building:

• SOC 2 Type II: we don't have a report yet. Some of the controls it tests already run and keep records: reviewed changes, secret scans and dependency checks. Ask us about our plans for it.

Your code is kept apart

Each customer has their own storage and their own encryption key protecting it, and their own model, which learns from that customer's work and never from another customer's. There is no pooled store of customer code, and no model that has learned from more than one customer's work.

If another customer's account were compromised, what would reach you? Nothing, by design: every request is checked against the account it belongs to, so their access stops at their own storage and their own model.

Encryption and connections

Your code and your model are encrypted where they are stored, under a key belonging to your account alone. Everything travels over an encrypted connection between you and us, and between our own services. We hold the keys for you today; customers who need to hold their own keys can ask about it as part of an enterprise agreement.

What Ryo learns from

Ryo learns each night from the work your team has approved. Today that means pull requests that were merged and whose checks passed. Work that was closed unmerged, or merged with failing checks, is not learned from. Ryo skips nights when little or nothing has been added or changed, like over a quiet weekend.

You choose which repositories Ryo may see, and which paths inside them to leave out. Secrets files are always skipped. If our filters spot a key, a password or personal details like ID numbers anywhere else, we remove the lines that hold them and show you each one when you set up. If we can't tell which lines a secret spans, we keep that file out and tell you why. Removing context for your model impacts performance and can keep it from giving you and your AI tooling the answers it needs.

Every settings change is logged with who made it and when, and so is each addition or removal of repositories — how many, never which. Changes made in your own repository are recorded in its git history.

What we never do

We do not train Airgetlam's own models on your data. Not on any plan, and not as a default you would have to find and switch off. Your team's work trains your Ryo model and no other.

Ryo runs on our own infrastructure. Your code reaches another company's AI service only as answers to the AI tools your team connects, under your company's own terms with their makers. We do not sell your data, and we do not use it for advertising.

We are constantly improving Ryo. To check that an update works, we compare counts and rates across customers, and we may test it on a copy of your own Ryo, in your own storage. That copy never posts or answers, and only its scores leave. No other customer's model, and none of ours, learns from your code.

Who can reach it

Two-step sign-in is required on every account that can change our code, our cloud or this website. Access is limited to the people who need it to run the service, and we review it every quarter.

Every change to the code that runs our service goes through a reviewed pull request: an automated check must pass, and the account that pushes a change can't approve it. The one exception is a written emergency procedure, and each use is logged and reviewed afterwards. A scan for credentials runs before any change can merge, and again across the full history every week. Only authorized personnel can approve changes or access production, and customers can request the current list.

Keeping it, and deleting it

Your code and your model are kept while you are a customer. You can disconnect a repository at any time, and you can ask us to delete everything. When you do, we start deleting at once: your code, your model and its search index, then every backup within 30 days. We email your account's contact when it's done. We keep only records that hold none of your code: billing records as long as tax law requires, and, after your account closes, what was agreed for 4 years, and size counts and an account history that names no one for 13 months.

Disconnect a repository and Ryo stops learning from it. We delete its code from your storage right away; what Ryo already learned stays with your model until you ask us to remove it or delete your account. If our GitHub app is uninstalled, we keep everything for 7 days in case that was a mistake, then delete it.

If something goes wrong

If customer data is exposed, we will tell the customers affected without undue delay and in any case within 72 hours of becoming aware, with what we know, what we are doing and what we recommend you do. If law enforcement asks us to wait, we wait only as long as it asks.

If you've found a security problem in our service, email security@airgetlamlabs.com. We aim to acknowledge it within two business days. If you act in good faith — testing only your own account, never reaching, keeping or changing anyone else's data, not degrading the service, asking for nothing in return, and giving us a reasonable chance to fix it before telling anyone else — we won't take legal action against you for that research. This covers our own systems only, not our providers'.

What we will send you

The sub-processor list is published alongside this page and needs nothing from you. Under a mutual non-disclosure agreement, we'll send a detailed description of how customer code and models are kept apart, and answer your own security questionnaire.

See our sub-processors