Laurie Kirk, described by Windows Latest as a Google researcher who joined Google in 2024 after four years as a Microsoft reverse engineer, argues that the Windows NT kernel handles one hard problem better than Linux: deciding and bounding what a program is allowed to do. Her reasoning is that NT’s object-based design and per-object security controls give a clearer answer to a question that AI agents make urgent, namely what exactly an agent is allowed to do. Her phrase “puts Linux to shame” is an opinion. No published, controlled cross-platform measurement supports it. She also imagines an alternate history in which Microsoft opened NT’s source early enough for a company such as an early Amazon to build its own variant.
What Kirk actually said
Kirk’s own words appear in a public LinkedIn post. Windows Latest, which published an interpretive article on September 26, 2026, built its coverage around that post and added its own framing about AI agents. Keep the two apart: the quotations below are Kirk’s, while the article’s synthesis is the publication’s.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Inside Windows Nt | $41.52 | Buy on Amazon |
| 2 |
|
Showstopper!: The Breakneck Race to Create Windows NT and the Next Generation at Microsoft | $16.47 | Buy on Amazon |
| 3 |
|
Windows Nt Shell Scripting | $27.20 | Buy on Amazon |
| 4 |
|
The Windows NT Device Driver Book: A Guide for Programmers | $14.00 | Buy on Amazon |
| 5 |
|
Windows NT File System Internals: A Developer's Guide | $37.16 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
“Say what you will about Windows (OS), but the NT kernel really is an engineering marvel that still puts Linux to shame in many ways.”
Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.
She also describes NT as “more like an object-oriented language, with a strong security model from day one,” and says she likes to imagine an alternate history where NT won. These are design judgments. They are not the output of a test, and they should be read that way.
#1 Best Overall
The historical backdrop is simple. Windows NT first shipped in 1993, according to the historical reference Windows Latest links, and later mainstream Windows releases were built on the NT line. The argument is therefore about the kernel architecture underneath Windows, not about the Windows 11 interface or the applications bundled with it.
What “object-oriented” means for NT
Kirk uses the term informally. Microsoft’s own documentation calls the NT design object-based rather than object-oriented in the classical C++ sense, and that distinction matters for reading her comparison precisely.
Microsoft’s Object Manager documentation on Microsoft Learn, dated May 12, 2025, says kernel resources such as files, devices, synchronization mechanisms and registry keys are represented as objects. It lists more than 25 object types and describes the manager’s work: creating and destroying objects, maintaining a namespace, tracking the resources each process holds, and tracking the access rights specific to each object.
Rank #2
Microsoft’s object-based design material adds a structural point. Executive components define object types and expose support routines for manipulating them. Other components call those routines rather than reaching into an object’s internals, which Microsoft says allows an implementation to change internally while the routines other components depend on remain stable. That is a claim about encapsulation. It does not show that every NT component is cleanly separated, and it says nothing by itself about how well Linux abstracts its own subsystems.
How NT’s security model is built
Microsoft’s security documentation describes a model based mainly on per-object rights, with some system-wide privileges alongside. Four elements do most of the work:
- Access tokens carry a process’s security context.
- Security descriptors can attach an access control list (ACL) to an object.
- Access checks compare the caller’s token with the object’s permissions.
- Handles store the rights granted when they were opened, and those rights are checked again on later I/O requests.
The handle rule matters for agents because it ties authority to the specific object a process actually holds. Two caveats apply. The model is not purely per-object, since system-wide privileges still exist. And real-world security also depends on driver behavior, default settings, and compatibility requirements that have accumulated over many years. Windows Latest flags those factors as shaping outcomes independently of the kernel’s design.
Rank #3
How Linux divides the same jobs
Kirk’s critique is that Linux’s permission and isolation controls can feel like overlapping pieces rather than one coherent authority model. She lists UIDs, GIDs, ACLs, cgroups, policies, SELinux and filesystem modes, and argues that NT could use explicitly typed resources and centralized audit trails. Windows Latest reads the Linux side differently: each mechanism was built to do a different job. The table places those jobs next to the NT mechanisms that Microsoft’s security documentation describes.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →| Job | Linux mechanisms described | NT mechanisms in Microsoft’s security documentation |
|---|---|---|
| Identify the principal | User IDs (UIDs) and group IDs (GIDs) | Access tokens carrying the process’s security context |
| Control access to an object | File modes and ACLs | Security descriptors with ACLs, checked against the caller’s token |
| Divide privileged power | Capabilities that split root’s powers | System-wide privileges alongside per-object rights |
| Give a process its own view of resources | Namespaces | Not stated |
| Limit resource use | Control groups (cgroups) | Not stated |
| Add policy enforcement | Linux Security Modules (LSM hooks; SELinux is one policy module) | Not stated |
| Restrict a process without extra privilege | Landlock, applied by a process to itself and passed to its children | Not stated |
| Keep rights attached to an open handle | Not stated | Granted rights stored with the I/O handle and checked on later requests |
“Not stated” means the Linux or Microsoft documentation consulted for this comparison does not describe that mechanism in these terms. It does not mean the platform lacks the capability.
The table shows that the two designs answer the question differently. Linux layers independent mechanisms, and part of that modularity is deliberate. The LSM framework provides hooks so that different policy modules can plug in, and Landlock lets an unprivileged process restrict itself and pass those restrictions to its children. Both are relevant to confining a program such as an agent. They also sit across several interfaces, which is the coherence problem Kirk is pointing at.
Why agent permissions reopened the old question
Kirk’s sharpest line is a question rather than a verdict: “What exactly is this AI agent allowed to do?” An agent often acts for a user while calling tools, reading files and changing system state, so the permission answer has to hold across all of those actions. Her broader prompt asks what an operating system kernel would look like if it were designed from scratch in 2026 with an eye to the future.
Microsoft has its own work in this area. At Build 2026, according to Windows Latest, Microsoft announced an early preview of Microsoft Execution Containers (MXC). In that model, developers declare the access an agent needs and Windows enforces the declaration at runtime. Check Microsoft’s Windows Developer Blog for current details and release status. The preview should not be read as general availability.
Whichever design a reader prefers, a fair comparison of operating systems for agents should ask:
Best Value
- How finely can authority be scoped, down to a single object or resource?
- How easily can a policy be composed, read and checked by someone who did not write it?
- Which actions leave an audit record, and where is that record kept?
- What compatibility constraints do existing applications impose?
- How much systems expertise does a correct configuration require?
- Who maintains the controls over the years, and how do updates reach them?
Current public documentation does not settle these six questions for either kernel.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.The Open NT alternate history
What Kirk proposes
Kirk says Microsoft “should have made an ‘Open NT’” in the early 2000s. She does not mean a fully GPL-style license. She means enough source access that a large organization could replace a memory allocator, customize a scheduler or network stack, while Microsoft kept control of a security and compatibility baseline. Her example is an early Amazon creating an “AmazonNT” for EC2. The scenario is explicitly counterfactual: it asks what might have happened, not what did.
What Microsoft already offered
Microsoft did run source-access programs. Windows Latest names Shared Source and the Windows Research Kernel, which is intended for academic teaching and research. As the article describes them, neither gave Amazon a general right to ship a commercial NT fork. Inspecting or studying source is not the same as a supported model for building a product on a modified kernel, and that gap is the real distance between Kirk’s scenario and the history that happened.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Why forks are expensive
The strongest counterargument concerns maintenance. Windows Latest quotes Linux kernel developer David Airlie on the cost of forks: a divergent kernel must keep integrating and testing upstream fixes as they arrive. Linux’s development model is built so that organizations can send their changes back into a shared project, which spreads that cost across contributors. An Open NT would have had to answer the same questions: who decides which changes flow back, and who pays for the integration work that keeps a fork current. The alternate history cannot answer them.
Other systems that bear on the question
Capability-based security is not unique to NT or Linux. Windows Latest points to FreeBSD jails and Capsicum as isolation and capability-oriented controls. They are useful if you are asking what a new operating system designed for agents might borrow, because they offer different answers to the same question: what is this process allowed to do?
How to weigh the claim
Kirk’s comparison can be argued on design grounds, but a few kinds of evidence would settle parts of it:
Quick Recap
- Controlled comparisons that run the same agent workload on each system and record what the agent could reach, with policies written by people of comparable skill.
- Published policy definitions for the same task on each platform, so that the expressiveness and length of the rules can be compared directly.
- Audit output from each system during an incident-style exercise, showing which actions were recorded and how easily an investigator could reconstruct them.
- Evidence from MXC, or any similar runtime enforcement layer, once it reaches general release, including how developers actually declare access.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
