Security
Isolated by default
What separates one workspace from another, what we store, and how to tell us about a problem.
Isolation
The boundary is the sandbox. Everything a run does happens inside it.
A machine per run
Its own filesystem and network namespace. One workspace cannot read another's files.
Loopback is the sandbox
A service a run starts is reachable inside the sandbox, not on your host.
Preview URLs are signed
A running service is exposed through a signed, expiring URL — never an open port.
Destruction is gated
A destructive action asks for approval regardless of your autonomy level.
Credentials
You connect them; we hold them encrypted and never display them again.
Encrypted at rest
Connection secrets are sealed with AES-256-GCM before they are stored.
Write-only in the UI
A stored secret can be replaced or revoked, never read back — including by us.
Scoped per workspace
A credential belongs to the workspace that connected it, not to the whole account.
Redacted in transcripts
A token pasted into a message is scrubbed before it reaches the event log.
What we store
Your work is yours. This is the whole list.
The transcript
Goals, tool calls, results and the run's own reasoning, so a thread can be replayed.
Artifacts you export
Stored in durable storage, so a deliverable outlives the machine that made it.
Usage records
Token and time accounting, so a bill can be explained line by line.
Reporting a problem
If you believe you have found a vulnerability, tell us before disclosing it publicly and give us a reasonable window to fix it. Include what you did, what you expected, and what happened — a reproduction is worth more than a severity rating.
We will confirm we received it, tell you whether we could reproduce it, and keep you updated until it is closed. We do not pursue researchers acting in good faith.
Read how the loop works
The documentation covers isolation, verification and the controls you keep.

