How we handle your access and your data.
This is how every engagement runs: what we ask for and why, where your information lives, how the sandbox is sealed and how we delete it all at the end.
What access we ask for, and why
We ask in this order, and each step is yours to grant. You set up every access we use, and you can revoke it at any time.
- A chat channel with your engineers. Ideally a guest account in your own Slack or Teams, so every message stays in your workspace under your retention rules, and we keep no copy once you remove us.
- One named contact and a kickoff call. You walk us through your architecture and who owns what. We agree a cap on your team’s time.
- Documents you already have. Architecture diagrams, service list, API specs, your last pentest report, onboarding docs. We read before we ask.
- Read-only access to your infrastructure code. Terraform, Kubernetes manifests, docker-compose and similar: the blueprint for the copy.
- View-only access to your cloud account. Recommended. Production drifts from the code, and that gap is where problems usually are. Configuration and metadata only, for the survey only, logged in your own audit log.
- Source code. Recommended, not required. With it, our agents read the code for flaws, then prove them against the copy. Without it, they attack as an outsider would.
What we never touch
- Production data or patient records
- The contents of your databases or storage. On AWS we use ViewOnlyAccess or SecurityAudit, never ReadOnlyAccess, which can read them
- Write access to your existing systems
- Production credentials
Prefer not to grant access yet?
We can run the same survey over screen-shared calls where you drive. It’s slower, and takes more of your team’s time.
Where your information lives
On a cloud workstation in your account
Our working copy of your code, configs and notes lives on a machine with an encrypted disk, inside the throwaway cloud account made for you. The agents can’t reach it. It has no ports open to the internet, we reach it through your cloud provider’s secure session service, and it shuts down when idle.
Not on a laptop
The engineer’s laptop is only a screen and keyboard into that workstation. Your files aren’t stored on it. It has full-disk encryption, auto-lock, remote wipe and automatic updates, and a separate browser profile for you, deleted at the end.
Short-lived access
Two-factor login (passkey or security key) on every account that touches your systems. Sign-on roles that expire. No long-lived keys on disk.
Chat is for talking
Files, exports, diagrams and credentials go into a storage bucket in your sandbox account, never into chat. If a secret lands in chat, we ask you to delete and rotate it. We use chat in the browser and don’t download attachments.
Your material stays out of backups and sync, out of any shared secrets store (you get your own), and out of every AI tool except the model provider we name to you. We sign your NDA, or bring a mutual one.
How the sandbox is sealed
- Its own throwaway account. By default a cloud account you own, created for this and deleted afterwards. Hosted by us if you prefer. No link to production, no shared credentials.
- No internet. No route out, outbound traffic denied by default, and DNS limited to an internal resolver. The agents sit in their own network that only reaches the copy.
- Public made private. Your gateways, load balancers and hostnames work as in production, but only inside the sandbox.
- Nothing from outside. Login, payment and email services are replaced by stand-ins. Tools are installed beforehand, so nothing is downloaded during a run.
- Nothing real inside. Synthetic data and freshly generated secrets, so even a failure would leak nothing real.
- Checked every run. Before each run we test that the internet is unreachable. All traffic is logged, outbound attempts raise alerts, and there is a kill switch.
What leaves the sandbox
We don’t run our own AI models. The agents reach a hosted model through one exit that allows only that model and is logged. What the agents read is sent to the model provider, so we tell you which provider we use before any run, and only use providers that don’t train on your data. We tell you how long each one keeps it.
What the sandbox doesn’t copy
Web firewalls, CDNs and bot filtering usually aren’t reproduced, so the test is harsher than a real attack. A finding your firewall may block today is labelled that way, and is still worth fixing. Flaws in that edge layer itself won’t be found. What’s exposed to the internet we check in the survey, from the view-only look at your real account. A few services have no private mode and get a small substitution.
How we delete everything
- At the end of the engagement we delete the workstation’s encrypted disk, your secrets store and the throwaway account itself
- We delete your browser profile on the engineer’s laptop
- You remove our access, and we confirm the deletion to you in writing
Questions about your setup?
Ask them on a free scoping call. We’ll walk you through how this works for your systems.
Request a free scoping call