Blog · 9 November 2026

Agent computers: a clean box is not a clean identity

Moving agents off employee desktops helps. It does not answer who the agent is, what it can reach, or whether your Conditional Access policies even see the session.

By Marcus Bengtsson, CTO at Sciber

Agent computers are arriving as a new kind of endpoint.

Not as another laptop to enrol. As a cloud box that signs into your mail, tickets, and cloud console as someone you already trust.

A CISO told Brad Menezes that desktops are dirty. I get the instinct. AI coding tools landed on employee laptops. Then a poisoned LiteLLM release stole credentials from wherever Python started. Days later, a tampered Claude Code extension in Cursor's marketplace tried to plant an SSH backdoor. Brad's post captures what many security leaders are saying in private: move development and agent execution off the laptop.

So move the agent. Problem solved?

Not quite. This article ties together three notes I published last week: where the agent runs, what it can reach, and which trust decision greets it when it signs in.

1. Where it runs is not who it is

A clean sandbox helps with one thing. It cannot read the laptop's cookies, SSH keys, or local files.

Too often the sandbox still gets the developer's own GitHub token and cloud credentials. Then the same reach just runs on someone else's computer.

The real question is not where the agent runs. It is who it is when it acts.

If the logs still show the person did what their agent did, accountability breaks. Incident response cannot tell the two apart. Lightspeed's recent CISO survey (n=200) found that only 20% have a full inventory of their agents, and 26% use or would use human credentials for them (Lightspeed Cyber 60).

Moving execution helps only when the agent also gets its own identity and a reach limited to the task. I wrote about that identity gap earlier; the Agent ID direction from Microsoft is useful here, but only if the agent you care about is actually covered.

2. The risk sits in the connections

A sandbox that cannot reach mail, tickets, or the IdP is useless. That is the quiet trade-off in every agent computer.

Vendors isolate the box so it cannot read the laptop. Then they connect the browser, the connectors, and the tokens that make the agent useful. OpenAI's sandbox security guidance states it plainly: agent-generated code can access the files, credentials, and network available to its environment. That is not a bug. It is the product.

Grok Bot's security FAQ makes a related point: the bot has no identity of its own, and blocking an OAuth connector does not block that service's website in the same browser. Anthropic's computer-use guidance says to isolate the VM, allowlist the internet, and avoid handing over login secrets, even though the tool exists to use the desktop and the web.

So the risk did not leave the desktop. It moved into the connections.

If the agent still signs in as the person, the sandbox inherits their reach. Isolation without a separate identity and a narrow allowlist is just a cleaner place to hold the same keys.

3. Linux, Conditional Access, and a new implicit trust

Many agent computers are Linux-class machines the company never enrolled. Grok Bot runs in a Firecracker microVM. Manus documents an Ubuntu cloud computer. OpenAI says Dot's cloud computer does not inherit the member's device policies.

That matters for Entra Conditional Access.

Linux is a supported platform in Conditional Access. The gap is not that CA cannot cover it. It is that many policies were built for Windows, macOS, and mobile, or accept compliant device OR MFA. Microsoft also notes that the platform signal comes from the user-agent string, which is not verified, and that agent sessions in cloud infrastructure may not present device compliance signals at all.

Intune on Linux is narrow: Ubuntu Desktop 24.04 or 26.04, or RHEL 9 or 10, with GNOME, Edge, and the Intune app. A vendor-hosted agent VM is not that path.

So a helpful agent can open a browser session as the user, pass MFA, and hold SaaS reach on a host you never managed. That is a new implicit trust. Microsoft already ships a template to block unknown or unsupported device platforms. Whether your tenant uses it, and whether Linux is intentionally managed or silently left out, is a tenant question, not a slogan.

What to check this week

None of this needs a new product first. It needs an honest look at how your policies treat Linux and sessions with no device state. Start in report-only. Exclude break-glass accounts. Validate in sign-in logs before you turn anything On.

1. Hunt. Look for successful browser sign-ins where the OS is Linux, or where deviceId is empty, against apps that hold company data.

2. Block unknown platforms, in report-only first. Use Microsoft's unknown or unsupported platform template. Add Linux to the exclude list only if you actually manage it.

3. Find every grant that is "compliant device OR MFA". For high-value apps, that OR is often how an unmanaged agent browser gets in. The alternate template exists for a reason; treat it as a pattern to hunt, not a default for apps that hold company data.

4. Tighten unmanaged browser sessions. Raise sign-in frequency. Set persistent browser sessions to never for non-compliant devices (Microsoft's ready policy). Where you have the licence, route high-risk apps through Defender for Cloud Apps session control.

5. Turn on Token Protection where it is supported. Linux is not on its supported list today. Pair it with the platform and compliance policies above so Linux is not the bypass story.

6. Prefer a separate agent identity over "sign in as the user". For classic apps that means a workload identity or service principal. Microsoft Entra Agent ID is the direction for agents Microsoft can see; check what is generally available in your tenant before you promise it, because some of the admin experience has still been in preview. Third-party agent computers that sign in as the human are a different subject. They still need their own scoped identity if you want IR to tell person and agent apart.

7. Broker short-lived tokens. No standing admin. Issue credentials at run time through a vault or secrets broker, scoped to the task.

8. Default-deny egress. Pin a package mirror. Allowlist hosts and extensions. A sandbox that can reach the whole internet with the user's cookies is not a containment story.

Closing

I am not arguing against agent computers. I use them. The USP of a dedicated Linux box is real: isolation from the laptop, a clean filesystem, room to work without touching the employee's local secrets.

The USP is also the risk. A useful agent must connect. Those connections land in SaaS through browser sessions and OAuth. If your Conditional Access story still treats device trust as a Windows and Mac problem, you have created a new implicit trust without meaning to.

At Sciber we spend a lot of time on identity, Conditional Access, and the boring controls that make agent reach auditable. The checklist above is the place I would start in any tenant this week.

If an agent signed into M365 from a computer you never enrolled, which of your policies would have evaluated it, and who would the logs say did the work?