Trust and security
What we protect, and what we have not built yet.
Rudran runs code that a language model wrote, on a machine with real tools. That is only reasonable if the isolation and the stopping points are real. Here is what they are.
Isolation
One sandbox per agent
Every agent runs in its own container with its own filesystem, shell and process space. One agent cannot see or reach another agent's machine.
Workspace-scoped data
Agents, threads, documents and usage are scoped to your workspace. Access is checked on every read and write, not only at sign-in.
Sandboxes hibernate
An idle agent is suspended rather than left running, which shrinks both the cost and the exposed surface.
Egress policy is built, not yet enforced
An allowlisted, CONNECT-only filtering proxy with SNI pinning is built and tested. It is not switched on in production yet, so today a sandbox is not network-restricted. We would rather say that here than let you assume otherwise; this line changes when enforcement does.
Human control
The approval gate is enforced on our servers, not by asking the model nicely. A sensitive action does not execute until a human decision is recorded against it, and the gate fails closed — if the platform is degraded, the action does not slip through.
Underneath that sits a hard safety floor: a class of destructive commands is refused outright regardless of what the agent, its playbook, or anything it read on the web asks for. It has fired on the live deployment.
Supervision is set per agent, so an agent doing research and an agent touching your systems do not have to share a policy.
Your data
Export
Export your data from settings at any time.
Account deletion
Delete your account from settings. It is removed from service immediately; the final erasure of the underlying records is still a manual step for us while we are pre-launch, so ask and we will confirm when it is done.
Threads are archived, not deleted
Archiving keeps the record of what an agent did rather than destroying it. You can always retrieve an archived thread.
Model providers see your prompts
The model you choose for an agent receives that agent’s context in order to answer. Choosing the model means choosing who processes that content.
Reporting a vulnerability
If you have found a vulnerability, the reporting address is at the bottom of this page. What follows is what you can expect from us, and what we ask of you in return.
What to expect. We aim to acknowledge every report within three business days, and to tell you our assessment — whether we can reproduce it, and how serious we think it is — within ten business days. If a fix needs longer, we will say so and keep you updated rather than go quiet. We do not operate a 24/7 on-call rotation, so please do not rely on an out-of-hours response.
Safe harbour. If you make a good-faith effort to follow this policy, we will not pursue or support legal action against you for your research, and we will treat your activity as authorised. If a third party brings action against you for research conducted in line with this policy, we will make it known that your activity was authorised.
Good faith means you access only data belonging to your own test accounts, you stop as soon as you have demonstrated an issue, you do not degrade the service for other people, you do not exfiltrate, alter or destroy data, and you give us a reasonable window to fix the issue before disclosing it publicly. Deliberately accessing another customer's data, denial-of-service attacks, and social engineering of our staff are outside this policy.
We do not yet publish a fix history with dated remediation times. We would rather say that than keep a page that quietly goes stale — when there is a real report to publish and someone to keep it current, it will appear here.
What we do not have yet
We have no SOC 2, ISO 27001, or any other compliance certification, and no audit is under way. When one starts, this page will say so with a date.
We do not publish an uptime figure or offer an uptime service level agreement. We would rather publish nothing than publish a number we cannot yet stand behind. System status is available if you want to watch it.
Audit-log export is built — owner and admin roles can export the org's audit trail as JSON or CSV. Single sign-on and customer-managed encryption keys are not. If your organisation needs them, tell us and we will tell you honestly where they sit.