Your provider keys
A key is encrypted before it is stored, with authenticated encryption under a secret that is kept outside the database. That secret can be rotated without asking you for your keys again.
A key lives with the address it opens. The service decrypts it only at the moment it sends a request to your endpoint, and it is never sent anywhere else.
Once a key is saved, neither the dashboard nor the API shows it again. They show its first five and last four characters so that you can tell your keys apart. You can replace or remove a key at any time.
Keys and tokens are kept out of logs and error messages. Tests in our codebase fail if one appears there.
Sandboxes
Every trial runs in a sandbox of its own, which is removed when the trial ends.
Your provider key never enters a sandbox. The agent reaches your model through Berkway, with a token that belongs to that one trial. Only a digest of the token is stored.
Berkway is the only path from a sandbox to your model. That is why a budget and the record of every call work the same way for every agent.
Accounts and sessions
Passwords are hashed with a slow, memory-hard function. A sign-in session is a random token in a cookie that scripts cannot read, that is sent only over HTTPS and that other sites cannot send for you. Only a digest of the token is stored, so a copy of the database cannot be used to sign in.
Every request is answered from the workspace of the person who is signed in. One workspace cannot read another’s connections, runs or traces.
Your runs
Traces and results are visible only to your workspace. We do not publish them and we do not train models on them. You can export a run, and you can delete it.
The one thing that can become public is the record of a run, and only when you publish it. A record states what was run and how it ended. It holds no trace, no task, no prompt, no model output, no key and no address of your own server. Each published record carries a digest of its content, and its page checks that digest in the reader’s browser.
A benchmark made of your own tasks is shown only to your workspace. To any other workspace it does not exist, and a published record of it numbers the tasks and does not name them.
Model output is treated as untrusted. Our code reads only answers it can validate, never prose.
Who else handles it
A few companies process information for us, under contracts that hold them to our instructions:
- A cloud provider that runs the sandboxes of runs. A sandbox receives the benchmark’s task and the commands the agent runs. It never receives your provider key.
- A cloud provider that hosts the service and its database.
- A host that serves this website.
- A mail provider that carries our email.
We give the current list, by name, to a customer who asks, and before a security review. Write to security@berkway.com.
What has not been checked
Berkway is a young company and the service is in pilot. It has not been audited against SOC 2 or ISO 27001, and it has not yet had an independent penetration test. This page will say so when that changes.
Until then, do not put regulated information into the service, such as health records or payment card numbers.
Report a vulnerability
Write to security@berkway.com. Say what you found, how to reproduce it and what it would let someone do. We will acknowledge your report within three working days and tell you what we do about it.
We will not take legal action over research done in good faith that keeps to these rules:
- use only accounts and data that are yours
- stop as soon as you can reach someone else’s data, and do not keep or share it
- do not degrade the service: no denial of service, no bulk mail, no social engineering of our staff
- give us a reasonable time to fix the weakness before you publish it
We do not run a paid bounty programme. We do thank reporters by name, if they wish. The same contact details are at /.well-known/security.txt for tools that look there.
