AI permissions: trust levels, allowed commands, root and the safety net
What an AI client may do on a device is set per device by an admin in the web dashboard (Devices → AI permissions), never by the AI. The Windows and macOS apps show the current state and open that page. Every change is audited, and every widening is e-mailed to all admins and owners.
Overview
| Setting | Values | Default | Widening needs |
|---|---|---|---|
| Trust | strict, standard, full | standard | full: your current password |
| Time limit for full | 1 hour, 4 hours, 24 hours, 7 days, or no limit | 1 hour | (part of full) |
| Run as | user, root | user | your password and the machine owner's consent (Linux only) |
| Safety net | on, off | on | off: your password |
| Safety net allows shutdown/reboot | on, off | off | on: your password |
Narrowing (back to standard, back to user, safety net on) needs no password.
Trust levels
- Strict: every high-risk call waits for a human approval; the allowed command list is not used.
- Standard: process_exec commands on the device's allowed list run without approval; every other high-risk call waits for an approval.
- Full: no per-call approval for any high-risk tool on the device, including Docker container actions and changes to scheduled tasks that run on it. WECOMMANDER's own containers are still never stopped, scheduled tasks on SSH servers still need an approval, and the safety net still applies. When a time-limited full trust expires, the device returns to standard.
Allowed commands
Under standard trust, routine commands can run without asking. Add patterns in the device's AI permissions (admin, browser session only). The rules are strict and fail closed:
- Only process_exec is ever allowed this way; tasks, terminals, deletes and kills always need approval.
- A pattern matches when its words are an exact prefix of the command's words: systemctl status matches systemctl status nginx but not systemctl stop nginx.
- A command containing ; & | $ ` < > ( ) { } [ ] \ * ? ! # ~, quotes, tabs or newlines is never allowed by a pattern, so df -h; rm -rf / cannot ride on df -h.
- Patterns cannot start with shells, sudo, su, env, xargs, interpreters, network or copy tools, database clients or destructive tools; multi-purpose tools such as docker, systemctl, git and apt need a subcommand.
- Each allowed run is audited as allowlisted with the pattern used.
Run as root: two keys (Linux)
- The machine owner allows it once on the machine: sudo bash /opt/wecommander/libexec/install-linux.sh --allow-root (undo with --disallow-root). The agent then reports that it can act as root.
- An admin turns run as root on for that device in the dashboard, with their password. Until the agent reports it can, the dashboard refuses and shows the command to run.
A web click alone can never give root on a server, and the agent enforces its half too: it refuses a root action without its own ceiling and a signed approval of that root action. As root, commands, background tasks and terminals run through sudo -n, and a file the service user cannot write is written as root (file tools still stay inside the allowed folders). Scheduled task runs stay the service user. Not available on Windows or macOS.
The safety net
On by default. It refuses a short list of catastrophic commands on every trust level and as root, even with a human approval:
- Recursive rm of /, /*, top-level system folders or the home folder, or with --no-preserve-root; recursive chmod, chown or chgrp of /.
- mkfs, mke2fs, mkswap, wipefs and blkdiscard on a /dev device; dd, shred or a redirection onto a whole disk.
- Fork bombs.
- shutdown, reboot, poweroff, halt, systemctl reboot or poweroff, init 0 or 6, unless “allow shutdown and reboot” is on.
- Stopping, disabling, masking or deleting the agent's own service, binary or sudoers rule.
It sees through sudo, env, nohup, timeout, sh -c, eval, $(…) and backticks, and lets ordinary work through (rm -rf ./build, systemctl restart). It is a guard against slips, not a sandbox.
Autonomous mode
Autonomous mode puts all devices, or the ones you select, on full trust for a time (5 minutes to 30 days) or until you stop it. It is started by an admin in the browser or the Management Center with their password; AI clients and API keys cannot start it. Every admin and owner is e-mailed, the dashboard shows a banner on every page while it is on, it stops automatically when the time ends, and stopping it needs no password. The device's own limits (allowed folders, read-only, no commands, secret masking) still apply.
SSH servers
Servers shared through Hybrid Access have the same settings on the Devices page under SSH servers: trust, time limit, run as, safety net and allowed commands. Only the owner of the computer they go through can change them. A server shared read-only refuses every change whatever its trust. Root needs the computer's sudo setting for that server to be sudoers as well.
- Approvals: web, e-mail and desktop notifications: How a high-risk AI action waits for a person, where approval requests arrive (dashboard, e-mail, Windows and macOS notifications, Linux Management Center) and what one approval allows.
- Agent: install, update and options: The WECOMMANDER agent on Windows, macOS and Linux: how it is installed and updated, the user it runs as, allowed folders, read-only and no-command switches, --allow-root, and checking its status.
- Hybrid Access: AI on SSH servers without an agent: Share saved SSH servers with AI clients through WECOMMANDER on your own computer (Windows, macOS, Linux): nothing installed on the server, keys never leave your computer, approvals and AI permissions apply.
- Security model in depth: The layers that keep AI access under control: outbound-only agent, allowed folders checked on the device, signed single-use approvals, safety net, secret masking, encrypted sync, OAuth and browser sessions.