Does an AI chat get full root on your server?
A connector acts with its own process rights. Our two-day root incident, what our connector allows now and a ten-minute server check.
🎯 For Server owners and developers connecting AI chats to their machines through MCP connectors.
- Understand that the connector's process rights are the chat's rights
- Separate client-side approval from server-side limits
- Check a server's connector in ten minutes
- Know what owner-signed root actions should look like
An AI chat gets exactly the rights of the process its connector runs as. If that process is root and keeps no command log, the chat can do anything root can, and afterwards nobody can say what it did. That happened on one of our servers for two days. This article explains what went wrong, what our connector allows today and how to check your own server before you connect an assistant to it.
The question in plain words
A connector is a small program on your server that an AI chat can call: list files, read a file, run a command. The chat does not log in to the server itself. It asks the connector, and the connector acts with its own operating-system rights.
So the useful question is not "which AI is safe". It is "which user runs the connector, which actions does it accept, and who writes down what it did". The chat's settings screen can ask you before each call, but that screen lives on the client side. The server has to enforce its own limits even when nobody is watching the chat.
What happened: two days of root without a record
From 31 October to 2 November 2026 a third-party remote-shell bridge ran on one of our hosts. It was installed for convenience: connect a chat to the server and let it work.
The bridge process ran as root. Its own usage counter showed 660 command executions. The server had no record of which commands those were. Access was revoked, the pairing state was removed, the remaining logs were kept as evidence and our own keys were rotated. The host held only our own keys; no client data, client credentials or client systems were involved.
We do not claim that nothing was read. With a root process and no command log, the absence of evidence is not evidence of absence. That is the honest limit of what we can say, and it is the reason the rules below exist.
The lesson is not about one vendor. The process had root rights and nothing narrowed them: no list of allowed actions, no separate user, no journal. Any bridge built that way carries the same risk.
What our connector allows today
After the incident we wrote our own connector, Tvijo Remote Desktop Commander, published at mcp.tvijo.com/mcp. It is the opposite design:
| Property | Third-party bridge (31.10–02.11) | Tvijo connector today |
|---|---|---|
| Process user | root | dedicated user remdesk, NoNewPrivileges, ProtectSystem=strict |
| Commands | any shell command | only named commands from a reviewed policy file |
| Arguments | anything | only where the policy declares a pattern |
| Files | the whole disk | only the workspace roots, symlinks resolved before the check |
| Record | none | every call appended to a hash-chained journal before the answer |
| Stop | uninstall | a stop file refuses every call on the next request |
| Sign-in | pairing code | OAuth with PKCE; the code is issued only after the operator enters a token on the consent page; access tokens live 15 minutes |
On 4 November 2026 we asked the live connector for its command catalogue. It returned 23 named commands: git operations limited to feat/remdesk-* branches, Go and Python test and build steps, read helpers such as ls, grep, head and wc, a pull-request helper for one repository and a queued coding task. There is no sh -c, no sudo and no free-form command.
What is still a risk
A short list of commands is not the same as no power. We name the remaining risks rather than hide them.
- Tests run code.
python.testandgo.testexecute code from the workspace, and the queued coding task can change that code. The real limit is therefore theremdeskuser and its sandbox, not the command names. The coding task itself runs in a separate unit that cannot see the connector's tokens or journal. Test commands, however, run inside the connector's own service, which can read its tokens and write its journal. Moving test runs into the separate unit is a fix we still owe. - Publishing is possible.
git.pushandpr.createcan send a branch and open a pull request. Branch names are restricted, and merging stays a human step. - The client approval is advisory. The chat app offers three settings per tool: allow always, ask every time, block. "Ask every time" is a good default, but it protects only against the chat, not against a stolen token. Server-side checks are the real boundary.
- One operator token. Today the consent page on
mcp.tvijo.comis confirmed with a single operator token. Per-user grants with scopes live onplatform.tvijo.com, and moving the connector onto them is planned, not finished.
How root work should look
Some tasks genuinely need root: restarting a service, renewing a certificate, reading a system log. We are building a separate executor for them, remdeskroot. It is not deployed. The design:
- The root process has no network. It listens only on a local socket. The internet-facing part runs as an ordinary user and has no shell.
- Actions come from a catalogue. Each action has a fixed argument list and a pattern for every parameter. The model chooses an action and parameters; it cannot write a command.
- The owner signs each risky action. The platform shows the owner what is asked, the owner approves, and the platform issues a ticket bound to this host, this action, these parameters and a short expiry. The ticket works once.
- Every call is journaled before the answer. The journal stores digests rather than file contents, and each line is chained to the previous one by hash, so a removed or edited line breaks the chain.
- One switch stops everything. A stop file refuses all actions, signed or not.
The approval table on the platform then becomes the request log: who asked, which parameter digest, what happened and which journal line proves it. A request without a journal line was not executed.
Check your own server in ten minutes
Before you connect any AI chat to a server, answer these questions on that server:
- Which user runs the connector?
ps -eo user,pid,argsand look for the connector process. If it is root, stop here. - Does it listen on the network directly?
ss -tlnpshows listening ports and their processes. A connector should sit behind a web server with TLS and authentication. - Can the model send a free-form command? Look at the tool list. A tool that takes a raw command string is a shell.
- Where are its limits written? You should find a policy file with allowed paths and commands. A promise in documentation is not a limit.
- Is there a journal, and can you read it? Run one harmless call and find it in the log with its time and decision.
- How do you stop it now? Know the command or file that refuses every call, and test it once.
- Which secrets live on this host? Provider keys, exchange keys and SSH keys should not sit on a host where an external agent connector runs.
The Model Context Protocol's own security best practices give the same direction from the protocol side: begin with minimal scopes, never pass client tokens through to other services, and sandbox anything that executes commands.
⚠️ When this does not work
A good connector cannot fix a server that already has a root bridge. If you find one, revoke access first, keep its state and logs as evidence, rotate the keys that were on the host and only then install a narrower connector.
A command list cannot tell a harmless test from one that a model has just rewritten to do something else. If the workspace holds anything valuable, the sandbox and the separate user matter more than the list.
An approval prompt that appears on every call trains people to click "allow" without reading. Keep prompts for actions that change something, and let read-only calls pass under the server's limits.
This article does not certify any product, including ours. It describes one incident, the connector we run now and a design that is still being built.
FAQ
Does an AI chat get root on my server when I add a connector?
It gets whatever rights the connector process has. If the connector runs as root without limits, then yes, in practice.
Is "ask before every call" in the chat app enough?
No. It is a useful extra check for the person at the screen. The server must still refuse anything outside its policy, because the chat app is not the only possible caller.
Was client data exposed in the October incident?
No client data, client credentials or client systems were on that host. Our own keys were, and we rotated them.
Can I use the Tvijo connector for my own server?
The current connector serves our own infrastructure. Customer access goes through grants on platform.tvijo.com with read-only scope. Root actions are not offered to anyone until the signed-ticket executor is deployed and reviewed.
Facts checked on 4 November 2026 against our incident record, deployment notes and the live connector's command catalogue. AI-assisted editorial text prepared for Tvijo. The root executor design is a plan, not a released feature.
Sources
Source links and observation dates are recorded below. Review status is stated for each material.
🔗 Article connections
Open the full knowledge graph ↗Published article links. Select a node to open its article at the heading.
Show the connection graph
Article links · keyboard / text view
Same topic
- Tvijo agent system: PRIZMA, RUNLEDGER and acceptance
- Read agent communication receipts without mistaking delivery for a human read
- Audit an AI technology radar against primary sources before promoting a story
- Run an AI coding harness evaluation offline before any live pilot
- DSHREM on a phone: SSH, tmux, and safe task handoff
- DSHREM owner practice: cwd, worktrees, WIP, and root boundaries
Next steps
Related reading
Tvijo agent system: PRIZMA, RUNLEDGER and acceptance
Tvijo is developing a controlled workflow from a repository task to an accepted result, with permissions, resource limits and attributable evidence. PRIZMA describes the intended resource profile; RUNLEDGER connects the work to its checks a…
AI and process foundationRead agent communication receipts without mistaking delivery for a human read
Find the current stage of a swarm message without confusing durable storage, delivery, runtime receipt and viewing. This guide connects local checks, the private observation screen and distinct receipts, so a delivery acknowledgement does not become an unsupported claim that an action completed.
AI and process foundationAudit an AI technology radar against primary sources before promoting a story
Check a technology news item against its primary source, version and event date before deciding to adopt it. This guide follows a fixed ten-item sample, separates facts from interpretation and forecasts, and preserves limitations and the condition for checking the claim again.
💬 Comments