Skip to content
AI Agent Security

Every AI agent now has its own computer. Who guards it?

6 min read

Between August and October 2026, four of the largest AI companies gave their personal agents a computer of their own. xAI's Grok Bot, Meta's Muse, OpenAI's dots and Anthropic's Claude Duet no longer borrow your screen for a minute: they work on a machine in the cloud, with a browser, files and the logins you connect, and they keep going while you are away. That moves the security question from what the model will say to what can happen on that machine.

What each company has published

AgentLaunchedThe machinePublished guard
xAI Grok Bot11 Aug 2026 (beta)One persistent cloud computer per account, shared by all of its BotsAuto Review with approval cards; local commands ask every time. The docs add: “Do not use separate Bots as a security boundary.”
Meta Muse8 Sep 2026Its own dedicated computer in the cloudA separate Sentinel agent on the same machine: “Nothing Muse does reaches the internet unless the Sentinel approves it.”
OpenAI dots29 Sep 2026Its own hosted computer and browser; your own machine can be linkedA read-only mode, and built-in rules for when to act and when to ask
Anthropic Claude Duet21 Oct 2026 (waitlist)Its own cloud computer, working while your laptop is offAsks before it sends messages, makes payments, signs or deletes anything

Sources: xAI's Grok Bot approvals, security and privacy page; Meta's Muse announcement; TechCrunch and Forbes on dots; R&D World on Claude Duet. All read on November 1, 2026.

What these guards have in common

Every published guard is a permission layer. It decides what the agent may ask for and when a person has to say yes: approve what leaves the machine, ask before acting, stay read-only while you are away. Those are the right controls for the conversation, and anyone running these agents should keep them on.

A permission layer judges the request. Once a request is approved (or covered by an “always allow” rule, which both Grok Bot and Muse offer), what happens next on the machine runs with the agent's rights. xAI says so in its own docs: an approval controls the proposed action and does not undo work already done. And when several agents share one computer, as Grok Bots on an account do, the docs are explicit that they are not a boundary from one another.

The layer below: what the machine itself allows

InnerWarden is that layer for the agents you run on your own Linux machines: Claude Code, Codex CLI, OpenClaw or anything else that runs commands and calls tools. It does not ask the agent what it intends. It decides on what the machine actually does:

  • The Execution Gate refuses, in the kernel, any program the agent was not authorised to start, scoped to the agent so the rest of the host carries on.
  • The Secret Read Guard refuses, in the kernel, reads of the secrets you protect.
  • The DNS Guard is a forwarding resolver that refuses to resolve malicious domains (on its threat-intel denylist, or rated high-confidence malicious by its classifier) while unknown domains still resolve. It runs in observe by default. The operator arms enforce; we recommend a rehearse first, and arming does not require one. No host we run has DNS Guard in enforce today: on our public challenge box it is set to observe, and only the challenge agent is pointed at it.
  • Per-agent identity (beta) ties each action to the agent instance that took it, so several agents on one machine stop being one blur.
  • Every allow, review and deny lands in a tamper-evident record you can check offline.

A hijacked agent can argue with a permission prompt. It cannot argue with the kernel. Since October 27 the AI agent on our live challenge is not even told it has a secret to protect; only the kernel stands between an attacker and that file. Anyone can try to get it.

Where this applies, and where it does not

The cloud computers in the table belong to their vendors, and we know only what they have published about them. InnerWarden runs on machines you control. That includes the server your coding agent works on, the host your OpenClaw runs on, and a Linux machine you link to an agent so it can reach your local files. Kernel enforcement is Linux only; on macOS and Windows InnerWarden monitors and screens. It is not an EDR replacement. The free Community tier screens commands and tool calls; the kernel layer is in Pro.

FAQ

Can InnerWarden protect Grok Bot, dots, Muse or Duet in their own clouds?

No. Those machines are run by their vendors. InnerWarden protects the Linux machines you run, including one you connect to an agent.

Should I turn off the agent's own approvals?

No. Keep them. Approvals decide what the agent may ask for; the kernel decides what the machine allows. The layers stack.

Does any of this need the model to cooperate?

The kernel controls do not. They hold outside the agent's process, whatever the agent was told. We set out how the layers combine in the four-layer hardening checklist.

Related reading

Keep following the attack path

Explore AI Agent Security
AI Security

AI Agent Hardening: A 4-Layer Checklist

AI agent hardening in four layers: the agent's settings, containment, a command and tool-call guard, and kernel enforcement on Linux. Who provides each.

9 min readRead
AI Security

Claude Code Hooks: Security Guide and Examples

Claude Code hooks for security: how a PreToolUse hook blocks a tool call, two working examples, what a hook cannot stop, and what the Linux kernel adds.

9 min readRead
AI Security

How Shell Rewrites Beat Command Blocklists

Attackers defeat text command filters by rewriting a command into a form that means the same thing but does not match the string. The fix is two layers: normalize before matching, and read the real argv in-kernel at execve.

8 min readRead
AI Security

Agentic AI Security Risks, Honestly Mapped

Most tools claim OWASP Agentic Top 10 coverage from the prompt layer. InnerWarden maps to all ten at the action layer with a reason chain, redaction, and a circuit breaker, and is honest about what is proven versus designed.

7 min readRead
AI Security

Indirect Prompt Injection: From Poisoned Input to Syscall

A poisoned input rewrites the agent's intent, and that intent lands on the host as execve, openat, or connect. Prompt-layer defenses live inside the context that just got poisoned. The damage becomes real at the syscall.

7 min readRead
Product

Protect AI Agents at Runtime, Down to the Kernel

Screen every command and tool call an agent makes on any OS; on supported Linux hosts, the Execution Gate refuses an unlisted program at exec.

Open