Security and control
AI should act on real machines only within limits you set, with sensitive steps approved by a person and every action recorded. This page describes how LetKite Lite does this today, and what the hosted LetKite plans add.
Authorization model
Client permissions
When you connect an AI client, you approve its permissions: fleet.read, fleet.exec and fleet.write. We recommend granting read and exec first, and write only when it is needed.
Machine permissions
Each machine has its own permissions, so a production database can be read-only while a test server accepts changes.
Account rights
By default, the gateway signs in to each machine with a dedicated account that has no sudo rights. Commands that need root must be listed explicitly in the sudoers file.
Sign-in with OAuth 2.1
AI clients connect to the gateway with OAuth 2.1, without a password or an SSH key:
- Access tokens expire after 15 minutes and are bound to your gateway's URL.
- Refresh tokens are replaced on every use. If an old one is used again, all tokens from that sign-in are revoked.
- Authorization uses PKCE (S256) and single-use codes that expire after two minutes.
- You approve each client on a consent page that shows its callback address. Only callback addresses on the allowlist are accepted.
- Admin passwords are stored as scrypt hashes. After five failed sign-ins from one address, further attempts from that address are blocked for increasing periods.
Confirmation for risky commands
Clearly destructive commands, such as deleting the root file system or formatting a disk, are rejected. Commands that are legitimate but risky, such as a reboot, stopping a service, clearing firewall rules or a recursive delete, are held until the AI restates the exact command and confirms it, so you can see what will run before it runs.
These checks are designed to catch mistakes, including a model misled by text it has read. They are not the security boundary: that is the account's own rights on each machine. For stricter control, run the gateway in read-only mode or do not grant fleet.write.
A record of every call
For each call, the gateway writes one line to its audit log and saves it to disk before returning the result: who made the call, when, on which machine, the command and its exit code. Output from long-running jobs stays on the machine and can be read later.
No software to install on your servers
Linux servers need no additional program. Enrollment creates the low-privilege account and adds the gateway's public key; it places no code and no private key on the server. Removing a server takes one command, and uninstalling the gateway first removes its key from every server.
Where your data goes
With LetKite Lite, the gateway runs on your own server. Commands and their output pass only between that gateway, your machines and the AI client you connected; LetKite itself does not receive them. The AI provider sees what its assistant reads and writes, under that provider's own terms.
Questions
Can the AI run any command?
No. A command runs only if the client's permissions, the machine's permissions and the rights of the account on that machine all allow it. Risky commands also need confirmation.
Does LetKite see my commands or files?
Not with LetKite Lite. The gateway runs on your own server; commands and their output pass only between that gateway, your machines and the AI client you connected.
Where is the audit log kept?
On the gateway, as one line per call. It records who made the call, when, on which machine, the command and its exit code.
How do I report a security issue?
Email contact@letkite.com. Please include steps to reproduce, and give us a chance to fix the issue before disclosing it.