Security and privacy

Devices connect out

• The agent opens the connection to the room over HTTPS and keeps it open. The room never connects in to a device, so a device needs no inbound port open to the network.
• Each device proves who it is with its own key pair and receives a short-lived token. An administrator can revoke a device at once, which stops its runs.
• A room can hold each new device for approval before it receives any work.

Signed releases

• Every release of the agent is signed with a key kept offline.
• Each release of the room pins one agent version. The room checks that version's signature and serves exactly that version to its devices.
• A device checks the signature again before it installs an update, installs only when idle, and goes back to the version it had if the new one does not start.
• The browser extension, Janus, is published through the Chrome Web Store.

The audit trail

• Every change to automations, files, runs, schedules, triggers, queues, devices, users, roles, the Vault and settings is written to the audit trail, with who did it and when.
• Sign-ins are recorded with the address they came from, failed ones too. Watching a run's screen live is recorded as well.
• Calls to the public API are recorded in their own ledger.

A room of your own

• Each client gets its own room: its own instance, with its own data and settings. Nothing is shared between rooms.
• Inside a room, permissions come from roles: Admin, Creator, Business, or roles you define. A draft is visible to its owner only.
• Passwords have at least 12 characters. A room can require an authenticator code or single sign-on (OIDC) for everyone. Repeated wrong passwords lock the account for a short time.
• A session ends when the browser closes.

Secrets

• Passwords and keys that automations need live in the Vault, encrypted. The room lists their names and never shows a value again. A run receives only the secrets its automation declares.
• A device's Windows credentials are write-only too, and stored encrypted for that device.
• A Credential's password is never written to a log.
• An API key is shown once, when it is made, and stored only as a hash.

What the AI sees

• Every AI call goes through the room. It is counted with its purpose, model, tokens and cost; the prompts and the answers are never stored. An administrator can set a monthly budget that stops calls at its limit.
• The KPI builder sends the model the outline of the steps, the names and types of the variables, and numbers, true or false values and counts from recent runs. A text value, a credential, a step's inputs and an error message never leave the room.
• Self-healing and AI steps work on the screen. To find an element again, or to act on an instruction, they send the model the window's element tree, the web page's labelled structure, or a screenshot. Screenshots can be turned off. A page's address is sent without its query or fragment.
• The room uses the AI provider it is set up with: OpenAI, Anthropic, or a compatible service.

Contact

Write to [email protected] with a security or privacy question, or to report a problem.