Your Claude Cowork session was humming along on a Tuesday morning. Commands were executing, files were being read, the terminal felt alive. Then Wednesday came, and suddenly nothing worked the same way. Local commands stopped responding. The agentic loop that had been so reliable the day before was now hitting a wall it couldn't see past. Chat still functioned. Basic file reads came back fine. But anything requiring the agent to reach into your local drive and act on it? Stopped cold.
If that happened to you around September 8th, you weren't dealing with a configuration mistake, a corrupted project, or a Claude bug. You were caught in the blast radius of KB5124008, a Windows 11 security update that quietly changed something deep in how Windows manages filesystem sharing with Linux-based virtual environments. This article unpacks exactly what changed, why it broke Claude Cowork and WSL, and what you need to do right now to restore full functionality.
The short answer: install KB5129195, Microsoft's remediation update. The longer answer is worth understanding, because the mechanism behind this failure reveals something important about how Claude Cowork actually runs on Windows, and that knowledge will serve you well the next time a patch cycle causes unexpected behavior.
The Fix First: How to Restore Claude Cowork After KB5124008
If your Claude Cowork stopped executing local commands after September 8th, the immediate fix is to install Microsoft's remediation update, KB5129195. This update moves affected machines to OS Build 26200.9457 (for devices on the latest release channel) or 26100.9457 (for devices on the 24H2 baseline). Once installed, Plan9 filesystem sharing is restored, and Claude Cowork should return to full agentic functionality.
Here is the exact path to apply the fix:
- Open Settings on your Windows 11 machine.
- Navigate to Windows Update.
- Click Check for updates.
- Let Windows download and install KB5129195 if it hasn't already applied automatically.
- Restart when prompted.
- After restart, press Win + R, type winver, and hit Enter.
- Confirm the build number reads 26200.9457 or 26100.9457.
That build number confirmation is important. Some machines that appeared to update successfully were still sitting on an intermediate build that hadn't fully applied the filesystem fix. Winver gives you the ground truth. If you see a build number lower than those listed above, run Windows Update again and check for any pending restarts or deferred updates blocking the install.
As PC Magazine confirmed in their coverage of the rollout, KB5129195 was Microsoft's direct response to the Plan9 sharing breakdown, and it specifically targets the Hyper-V virtualization layer where the regression was introduced. This isn't a workaround. It's a targeted fix from the team that introduced the problem.
If you're running an ARM64 device, note that the original breaking update was KB5124012 rather than KB5124008. The same remediation path applies: check Windows Update, install KB5129195, and verify the build number via winver.
What KB5124008 Actually Changed Under the Hood
KB5124008 (x64) and KB5124012 (ARM64) were Microsoft's September 2026 security updates, released on September 8th. On the surface, these were routine Patch Tuesday releases, bundling security hardening across multiple Windows subsystems. The change that caused the Claude Cowork disruption was not announced as a behavioral change. It was buried inside a modification to the Hyper-V Plan9 filesystem-sharing implementation.
To understand why that matters, it helps to understand what Plan9 is and why Windows uses it at all.
Plan9 and the 9P Protocol: The Hidden Bridge Between Windows and Linux
Plan9 (often referred to in technical contexts as the 9P protocol) is a network filesystem protocol originally developed at Bell Labs as part of the Plan 9 operating system. It provides a clean, message-based interface for sharing filesystems across process boundaries, including across virtual machine boundaries. Microsoft adopted 9P as the mechanism for letting Linux guest environments inside Windows Subsystem for Linux and Hyper-V-managed VMs access the Windows host filesystem.
When you run WSL and navigate to /mnt/c/ to access your Windows C drive from inside Linux, that's 9P doing the work. When Claude Cowork's Linux-based execution environment needs to read a file from your Windows project folder or write output back to your desktop, it's also going through 9P. The protocol acts as an interpreter between two completely different filesystem conventions, Windows NTFS on one side, and Linux-native filesystems on the other.
The 9P communication flow works through a series of typed messages. A client (in this case, the Linux VM) sends messages to a server (the Windows host) to negotiate connections and perform filesystem operations. One of those message types is called a Tattach request. Tattach is how the client establishes an authenticated connection to the filesystem it wants to access. It says, essentially: "I am this user, I have these credentials, please give me a file handle I can use to navigate this tree."
Token-Backed Tattach and the EINVAL Error
Modern Windows Hyper-V implementations use token-backed Tattach requests. Instead of passing plain credentials, the client passes a Windows security token, a structured object that encodes the user's identity, group memberships, and privilege levels. This approach is more secure than earlier credential-passing methods because it integrates with Windows' native security model and allows the host to make fine-grained authorization decisions before granting filesystem access.
The change introduced in KB5124008 caused the Hyper-V Plan9 server to reject these token-backed Tattach requests with an EINVAL error. EINVAL is a standard POSIX error code meaning "invalid argument." From the Linux VM's perspective, it sent a properly formatted Tattach request and received back a response saying the arguments were invalid. The VM couldn't establish the filesystem connection, which meant it couldn't mount the host drive.
As Windows Latest reported when Microsoft acknowledged the issue, this was a regression in the September update, meaning a change that worked correctly in previous builds introduced a defect that broke existing functionality. Microsoft's own acknowledgment confirmed that the Plan9 sharing implementation was the source of the problem.
The practical consequence was immediate and disorienting for anyone using Linux-backed tools on Windows. The Linux environment was running normally inside its VM. Windows itself was functioning. But the bridge between them was broken at the authentication handshake level, before any file data could even begin to flow.
Why Claude Cowork Was Specifically Affected
Claude Cowork runs its execution environment inside an HCS-managed virtual machine on Windows, which means it depends directly on Hyper-V Plan9 filesystem sharing to access your local files. HCS stands for Host Compute Service, Microsoft's API layer for managing lightweight virtual machines on Windows. When Claude Cowork needs to run a terminal command, read a local file, or write output to your project directory, it's doing that from inside a Linux environment that reaches across the 9P bridge to your Windows host.
This architecture is actually one of Claude Cowork's strengths under normal conditions. Running the execution environment in an isolated VM means that agentic operations are sandboxed, reducing the risk that a runaway command can damage the host system. The tradeoff is that Claude Cowork's local command capabilities are entirely dependent on that VM-to-host bridge remaining functional.
When KB5124008 broke the token-backed Tattach handshake, Claude Cowork's Linux VM could no longer mount the Windows host drive. The result was a partial failure that confused a lot of users: chat still worked, and basic file reading and editing often still worked, because those operations could sometimes fall back to alternative mechanisms or cached state. But anything requiring the agent to execute a command that touched the local filesystem, run a build process, invoke a script, or navigate the project tree in an agentic loop, failed.
The GitHub issue thread in the Anthropics Claude Code repository filled up quickly with reports from users across different hardware configurations describing the same pattern: agentic tasks stopping mid-execution, "cannot access path" errors appearing in contexts where the filesystem had been accessible minutes before the update, and WSL-based tools failing in parallel.
For teams using Claude Cowork as part of a real workflow, whether that's automating code reviews, running test suites, or managing internal tooling, this wasn't a minor inconvenience. It was a full stop on productivity for the subset of tasks that made the tool valuable in the first place. Understanding the architecture helps clarify why chat survived but command execution didn't: chat operates at the API layer, while local command execution requires the filesystem bridge that KB5124008 broke.
WSL Users Hit the Same Wall
Windows Subsystem for Linux uses the same Hyper-V Plan9 infrastructure that KB5124008 broke, which is why WSL failures appeared alongside Claude Cowork failures. The two tools share the same underlying architecture: a Linux VM managed by HCS, communicating with the Windows host filesystem over 9P.
If you use WSL for development, you likely saw one or more of these symptoms after September 8th:
- The
/mnt/cor other Windows drive mounts became inaccessible or threw permission errors. - WSL sessions that had been running before the update continued to function until restarted, after which they could not remount the Windows drive.
- Commands that write output to Windows-hosted directories failed silently or with cryptic error messages that didn't point to the real cause.
- Tools like Docker Desktop on WSL2, Git operations targeting Windows-hosted repositories, and Node or Python scripts accessing Windows project folders all behaved erratically.
The fact that WSL and Claude Cowork broke in exactly the same way is actually useful diagnostic information. It confirms that the failure was at the infrastructure layer rather than anything specific to either tool's implementation. If WSL mounts were broken after September 8th, Claude Cowork's local command execution would fail for the same reason, and vice versa. The fix is also the same: KB5129195 restores Plan9 sharing for both.
For development teams that rely on WSL for their daily workflow, the September 8th update was a significant disruption. The remediation update resolves it, but the experience highlights how deeply integrated the 9P bridge is into Windows' Linux interoperability story, and how a single regression in that layer can simultaneously break multiple tools that appear unrelated from the outside.
The Timeline: From Patch to Acknowledgment to Fix
The sequence of events between KB5124008's release and Microsoft's remediation provides useful context for understanding how these incidents unfold. Knowing the timeline helps you calibrate how quickly to expect resolution the next time a patch cycle causes unexpected behavior.
| Date | Event | Impact |
|---|---|---|
| September 8, 2026 | KB5124008 (x64) and KB5124012 (ARM64) released | Plan9 token-backed Tattach requests begin returning EINVAL |
| September 8–10, 2026 | User reports begin accumulating in GitHub issues and community forums | Claude Cowork agentic failures and WSL mount errors widely reported |
| September 12, 2026 | Microsoft acknowledges the regression publicly | Confirmed that Plan9 sharing was broken by the September update |
| Post-acknowledgment | KB5129195 shipped as out-of-band remediation update | Moves affected PCs to OS Build 26200.9457 or 26100.9457 |
The four-day window between the update release and Microsoft's public acknowledgment is relatively fast for enterprise software incidents of this nature. The disruption was widespread enough, and the affected tooling high-profile enough, that the signal-to-noise ratio in the bug reports was clear from early in the cycle. The GitHub issue thread at Anthropic's repository was a significant source of structured user reports that helped establish the pattern quickly.
One thing worth noting: Microsoft's original messaging around the September update had urged users not to delay installation, framing it as an important security release. That advice was technically correct from a security posture perspective, but it meant that many users who applied the update promptly were the first to experience the filesystem regression. This is a recurring tension in patch management: security urgency and stability risk sometimes point in opposite directions, and there's no clean resolution to that tension. The best response is to know how to diagnose and recover quickly, which is exactly what KB5129195 makes possible.
What This Means for How You Think About Claude Cowork on Windows
The KB5124008 incident is a useful prompt for understanding the infrastructure stack that Claude Cowork sits on top of when running on Windows, and for building habits that help you diagnose similar issues quickly in the future.
Claude Cowork is not a simple desktop application. On Windows, its agentic capabilities depend on a virtualization stack: HCS manages the VM, the 9P protocol handles filesystem sharing, and the Linux execution environment runs commands on your behalf inside that VM. That architecture delivers real security benefits, but it also means that changes to the Windows virtualization layer can affect Claude Cowork's behavior without any change to Claude Cowork itself.
Practically speaking, this means a few things for how you manage your setup:
Build Number Awareness Is a Useful Habit
Knowing your current Windows build number costs nothing and gives you immediate context when something stops working. After a Patch Tuesday release, checking winver before and after the update takes thirty seconds and gives you a clear reference point if you need to file a bug report or search for known issues. The build numbers 26200.9457 and 26100.9457 are now known-good references for the Plan9 fix. If you're below those numbers and experiencing filesystem issues, you have a clear path forward.
Partial Failures Are a Diagnostic Signal
The pattern that confused many users during this incident, chat works, file reading works, but command execution fails, is actually a useful diagnostic fingerprint. When Claude Cowork works partially but not fully, the failure is almost always at an infrastructure layer rather than in the AI itself. Full agentic command execution requires the filesystem bridge. If that bridge is broken, you'll see exactly this pattern. Future incidents with a similar signature should prompt you to check for recent Windows updates and look for known issues in the Hyper-V or WSL stack before troubleshooting Claude Cowork's configuration.
WSL Health Is a Proxy for Claude Cowork Health
Because WSL and Claude Cowork share the same 9P infrastructure, WSL's ability to mount Windows drives is a quick proxy test for whether the underlying virtualization layer is healthy. If ls /mnt/c returns your Windows files normally from inside a WSL session, the Plan9 bridge is working. If it throws an error, you know the issue is below the application layer. This diagnostic shortcut would have saved many users hours of troubleshooting during the September incident.
For teams where Claude Cowork is integrated into real workflows, those using it for marketing automation, internal tooling, or code generation workflows, understanding this architecture is part of being a responsible operator of the tool. If you're building workflows around Claude Cowork's local capabilities and want a structured framework for doing that reliably, the AdVenture Media guide on how founders are using Claude Code to build internal tools covers the architectural considerations that matter for production use.
Diagnostic Checklist: Is KB5124008 Still Affecting Your Machine?
If you're unsure whether your machine has been affected by the September update, or whether the remediation has fully applied, this checklist will give you a clear picture.
| Check | How to Run It | Healthy Result | Problem Signal |
|---|---|---|---|
| Build number | Win + R, type winver | 26200.9457 or 26100.9457 or higher ✅ | Build number lower than these values ❌ |
| WSL drive mount | Open WSL, run ls /mnt/c |
Windows files listed normally ✅ | Permission denied or "cannot access" error ❌ |
| Claude Cowork local command | Ask Claude to list files in a local directory | File listing returns successfully ✅ | Command fails or returns empty/error ❌ |
| Windows Update history | Settings > Windows Update > Update history | KB5129195 shows as successfully installed ✅ | KB5129195 absent or showing failed install ❌ |
| Architecture check | Settings > System > About, check System type | Identifies whether x64 or ARM64 ✅ | ARM64 users confirm KB5124012 was the original breaking update ⚠️ |
If any of the problem signals appear after you've applied KB5129195, the most likely culprits are a pending restart that hasn't completed the install, a Windows Update service issue that caused the download to fail silently, or a corporate Group Policy configuration that's deferring the update. In a managed enterprise environment, your IT administrator will need to push KB5129195 through your patch management system if automatic updates are controlled centrally.
The Broader Architecture: Why Windows Uses This Stack for Linux Interoperability
Understanding why Microsoft chose the Hyper-V Plan9 approach for Linux interoperability on Windows helps explain both the power and the fragility that the September incident revealed.
The alternative to running Linux in a VM would be to compile Linux-native applications directly for Windows, either natively or through a compatibility layer that translates Linux system calls to Windows equivalents. Microsoft actually tried an earlier version of this approach in WSL1, which used a translation layer rather than a full VM. WSL2 moved to the VM approach because running a real Linux kernel inside a lightweight VM delivers dramatically better compatibility. Real Linux applications that rely on kernel behavior they can predict simply work, rather than sometimes working and sometimes failing because the translation layer doesn't implement every system call correctly.
The tradeoff is that the VM approach requires a robust bridge for the host filesystem. If you run Linux in a fully isolated VM with no access to the Windows drive, the practical utility drops sharply. Most development workflows involve code that lives on the Windows side and needs to be accessible from the Linux environment. The 9P protocol solves this by providing a well-defined, message-based interface that both sides can implement reliably.
The security model around token-backed Tattach requests was a genuine improvement over earlier approaches. By using Windows security tokens rather than simpler credential passing, the implementation aligns with how Windows authorization works natively, and it allows the host to enforce consistent access controls. The September update introduced a defect in how those tokens were validated or processed on the host side, causing legitimate requests to be rejected. This was a regression in the implementation, not a fundamental problem with the design.
For professionals who use Claude Cowork as part of a real work setup, particularly those integrating it into marketing or development workflows, understanding that the tool runs on this kind of infrastructure is valuable context. It's why Claude Cowork on Windows has the security properties it does, and it's why changes to the Windows virtualization stack can affect the tool even when nothing in the tool itself changed. The AdVenture Media breakdown of Claude Code capabilities by role covers how these architectural realities translate into practical workflow considerations for marketing teams.
ARM64 Users: KB5124012 and What's Different for You
If you're running Windows 11 on an ARM64 device, such as a Snapdragon-based laptop or a Surface Pro X, the breaking update was KB5124012 rather than KB5124008, but the root cause and the fix are identical.
Microsoft distributes separate update packages for x64 and ARM64 architectures because the compiled code is different for each, but the logical changes packaged in a given Patch Tuesday release are typically the same across both. In this case, both KB5124008 and KB5124012 included the same modification to the Hyper-V Plan9 implementation, which means ARM64 users experienced exactly the same EINVAL rejection on token-backed Tattach requests and the same loss of filesystem access for Linux VMs.
The remediation, KB5129195, covers both architectures. When you run Windows Update on an ARM64 device, it will download the ARM64 version of the fix automatically. The target build numbers are the same: 26200.9457 or 26100.9457. Verification via winver works identically regardless of architecture.
ARM64 Windows adoption has been growing steadily, particularly among users who prioritize battery life and prefer the efficiency of ARM-based chips. If you're in this group, the KB5124008/KB5124012 incident may have been your first encounter with a platform-specific update causing a Linux interoperability regression. The experience is useful context: ARM64 Windows and x64 Windows share the same logical update cycle, but if you're troubleshooting a WSL or Claude Cowork issue, it's worth confirming which update package applies to your device before searching for solutions.
Preventing Future Disruption: A Monitoring Approach for Claude Cowork Users
The KB5124008 incident is unlikely to be the last time a Windows update affects Claude Cowork's local capabilities, because the tool depends on infrastructure that Windows actively maintains and modifies. Building a lightweight monitoring habit now will help you catch similar regressions faster in the future.
Track the Anthropic GitHub Repository for Claude Code
The GitHub issue tracker for Claude Code at github.com/anthropics/claude-code is where problems get reported and discussed in real time. Watching the repository (using GitHub's Watch feature, set to "Issues only") means you'll receive a notification when a new issue matching known patterns appears. During the September incident, the issue thread had substantial community activity within 24 hours of the update rolling out. If you had been watching the repo, you would have had a probable diagnosis before spending time troubleshooting your own setup.
Note Your Build Number Before and After Patch Tuesday
Patch Tuesday falls on the second Tuesday of each month. The day before and the day after each Patch Tuesday, spend thirty seconds with winver and note the build number. This habit costs almost nothing and creates a clear reference point if something breaks in the days following an update. You'll know immediately whether a behavioral change in Claude Cowork is correlated with a Windows update.
Keep a WSL Smoke Test in Your Toolkit
A simple WSL smoke test, opening a WSL terminal and running ls /mnt/c to confirm Windows drive access, takes under ten seconds and tells you immediately whether the Plan9 bridge is healthy. If you build this into your morning routine on Patch Tuesday weeks, you'll catch filesystem regressions before they affect your Claude Cowork sessions.
Understand Your Enterprise Update Policy
If you work in an organization where IT controls Windows Update policies, the timeline for receiving remediation updates like KB5129195 may be longer than for consumer devices. Enterprise environments often defer updates by days or weeks to allow for testing. If you're waiting on IT to push KB5129195, document your symptoms clearly, reference the known issue (Plan9 token-backed Tattach EINVAL, resolved by KB5129195), and provide your current build number from winver. This gives your IT team exactly what they need to prioritize and apply the fix.
For teams that have built structured Claude Cowork workflows into their operations, knowing when and how to escalate infrastructure issues is part of responsible deployment. If you're at the stage where you need a more structured approach to training your team on Claude Cowork and Claude Code, the AdVenture Media guide on when you need structured Claude Code training covers the signals that indicate it's time to move beyond ad-hoc learning.
Frequently Asked Questions
What is KB5124008 and why did it break Claude Cowork?
KB5124008 is a Windows 11 security update released on September 8, 2026, for x64 devices. It modified the Hyper-V Plan9 filesystem-sharing implementation in a way that caused token-backed Tattach requests to be rejected with an EINVAL error. Claude Cowork's Linux execution environment uses this Plan9 bridge to access the Windows host filesystem, so when the bridge broke, local command execution stopped working.
What is KB5124012 and how is it different from KB5124008?
KB5124012 is the ARM64 equivalent of KB5124008, released on the same date with the same logical changes. The two updates are separate packages because they compile for different processor architectures, but both contained the Plan9 regression. ARM64 users who experienced Claude Cowork failures after September 8th were affected by KB5124012.
What is the fix for the KB5124008 Claude Cowork problem?
Microsoft released KB5129195 as the remediation update. Go to Settings, then Windows Update, then Check for updates. Install KB5129195, restart your machine, and confirm the resulting build number is 26200.9457 or 26100.9457 using winver.
Did KB5124008 affect all Claude Cowork users on Windows?
The update affected Claude Cowork users running Windows 11 on machines where the update was applied. Users who had update installation deferred, either manually or through an enterprise policy, would not have experienced the issue until the update was applied. The failure was consistent across affected machines: partial functionality (chat, basic file reading) with loss of local command execution.
Why did chat still work in Claude Cowork if the update broke local commands?
Chat operates through Claude's API layer and does not require the Plan9 filesystem bridge. Local command execution requires the Linux VM to mount and access the Windows host filesystem, which depends on the Plan9 connection that KB5124008 broke. Operations that don't require that connection continued to function, which is why the failure appeared partial rather than total.
Was WSL also affected by KB5124008?
Yes. WSL uses the same Hyper-V Plan9 infrastructure as Claude Cowork's execution environment. Users whose WSL mounts stopped working after September 8th were experiencing the same EINVAL rejection on token-backed Tattach requests. KB5129195 resolves the issue for both WSL and Claude Cowork.
What is a Tattach request and why does it matter?
In the 9P filesystem protocol, a Tattach request is how a client (the Linux VM) establishes an authenticated connection to the filesystem it wants to access on the server side (the Windows host). KB5124008 introduced a defect that caused the host to reject these requests as invalid, preventing the VM from establishing any filesystem connection at all.
What does EINVAL mean in this context?
EINVAL is a standard POSIX error code meaning "invalid argument." When the Linux VM sent a token-backed Tattach request, the Windows host returned EINVAL, indicating it considered the request malformed or invalid. This prevented the VM from mounting the Windows filesystem, even though the request was formatted correctly according to the protocol.
How do I confirm KB5129195 was successfully installed?
Open winver (Win + R, type winver, press Enter) and confirm your build number reads 26200.9457 or 26100.9457 or higher. You can also check Settings, then Windows Update, then Update history, and look for KB5129195 listed as successfully installed.
What if my machine is still showing issues after installing KB5129195?
First, confirm the build number via winver. If it's still below the target builds, check for a pending restart that hasn't completed the install. In enterprise environments, contact your IT administrator to confirm KB5129195 has been pushed through your patch management system. If the build number is correct but issues persist, check WSL drive mounts independently as a proxy test for Plan9 health.
Does this issue affect Windows 10 users?
The verified facts about KB5124008 and KB5124012 are specific to Windows 11. The Plan9 implementation and HCS-managed VM architecture described in this article are Windows 11 features. Windows 10 users running Claude Cowork would not be affected by these specific updates.
Where can I follow ongoing developments about Claude Cowork on Windows?
The Anthropic Claude Code GitHub repository at github.com/anthropics/claude-code is the primary place where issues are tracked and discussed. The Windows Latest publication also provides timely coverage of Windows update regressions affecting developer tooling.
Key Takeaways
- The fix is KB5129195. Go to Settings, then Windows Update, then Check for updates. Confirm the resulting build is 26200.9457 or 26100.9457 via winver.
- The cause was a Plan9 regression in KB5124008 (x64) and KB5124012 (ARM64), released September 8, 2026. The update caused token-backed Tattach requests to be rejected with EINVAL, breaking the filesystem bridge between Linux VMs and the Windows host.
- Claude Cowork and WSL were both affected because both depend on the same Hyper-V Plan9 infrastructure. Chat worked; local command execution did not.
- Partial functionality is a diagnostic fingerprint. When chat works but commands don't, the failure is at the infrastructure layer, not in Claude itself.
- WSL drive mount health is a proxy test for whether the Plan9 bridge is working. If
ls /mnt/creturns your Windows files, the infrastructure is healthy. - ARM64 users were affected by KB5124012, not KB5124008. The root cause and fix are identical across both architectures.
- Enterprise users may need IT to push KB5129195 through a patch management system if automatic updates are centrally controlled.
- Watching the Anthropic Claude Code GitHub repository and noting build numbers around Patch Tuesday are low-effort habits that accelerate diagnosis when similar issues occur in the future.
If the September update disrupted how your team uses Claude Cowork, and you want to build workflows that are more resilient to infrastructure changes going forward, AdVenture Media runs live Claude Code workshops designed for professionals who use these tools in real work contexts. The sessions cover architecture, workflow design, and hands-on practice with the kinds of agentic capabilities that KB5124008 temporarily took offline. Find upcoming dates and register at adventuremedia.ai/workshops.
Join our live Claude Code workshops and training events
Learn more →
