Windows desktop 26.721.4979.0 exits when in-app browser GPU process loads vk_swiftshader.dll

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #35352
• #35411
• #35475
• #35505
• #35625

Powered by Codex Action

Additional confirmed reproduction on the same Windows desktop build.

Environment

• Windows 11 x64 (build 28000)
• Microsoft Store/MSIX package: OpenAI.Codex_26.721.4979.0_x64__2p2nqsd0c76g0
• Codex CLI: 0.145.0
• GPU: NVIDIA GeForce RTX 5070
• NVIDIA driver: 32.0.16.1062

Reproduction

1. Repair the app in Windows Settings.
2. Launch Codex desktop.
3. Resume an existing local session that contains in-app-browser state.
4. The app closes about 25 seconds after launch.
5. The package returns to Modified, NeedsRemediation; a later launch may show “This app can't open — Check the Store for more info about ChatGPT.”

The restored session is large (about 42.7 MB / 11,039 JSONL records), but it parses successfully with zero malformed records. It appears to be a reliable trigger for the browser/GPU restoration path rather than a corrupt session.

Correlated evidence

At the exact crash timestamp, the desktop log records:

Recoverable Chromium child process gone exitCode=101457950 processType=GPU reason=crashed serviceName=GPU
Recoverable Chromium child process gone exitCode=18 processType=GPU reason=launch-failed serviceName=GPU

Windows Microsoft-Windows-CodeIntegrity/Operational Event 3033 records that ChatGPT.exe attempted to load the packaged vk_swiftshader.dll, which did not meet Microsoft signing-level requirements. The DLL reports version 5.0.0 and Get-AuthenticodeSignature reports NotSigned.

A Crashpad dump (~45 MB) was created at the same timestamp. It is available privately if maintainers need it, but I have not attached it publicly because it may contain process memory or sensitive data.

Troubleshooting / comparison

• Repair temporarily clears the package status, but the same launch/resume sequence corrupts it again.
• Reinstalling or updating the Store app has not prevented recurrence (4–5 consecutive cycles).
• Disk space is sufficient, no reboot is pending, and Microsoft Defender reports no detection.
• The display driver reports healthy and there are no matching system display-driver reset events.
• The native CLI on the same machine is healthy: codex doctor --summary --ascii --no-color reports 17 ok | 1 idle | 0 warn | 0 fail.

This strongly matches the packaged Chromium/SwiftShader Code Integrity failure described in this issue. A safe workaround is to continue the same project through the CLI until the desktop package/browser GPU path is fixed.

confirmed, same issues:
Status: 0xC0000428
RequestedPolicy: 8
ValidatedPolicy: 1
SignatureType: 4
ValidatedSigningLevel: 6
VerificationError: 16

Get-AuthenticodeSignature reports the individual vk_swiftshader.dll as NotSigned, while the containing MSIX CodeIntegrity.cat signature reports Valid.

Windows AppModel-Runtime then recorded:

Event 6: process creation failed with 0x3CFC while checking machine-level package status
Event 217: destroyed the Codex Desktop AppX container

Crashpad generated a 40.9 MB dump at 09:50:00. The dump exception is 0x80000003 inside chrome.dll.

After Windows attempted automatic registration/ACL repairs, Get-AppxPackage OpenAI.Codex still reported:

Status: Modified, NeedsRemediation

Additional confirmed reproduction on Windows 10:

• Codex Desktop/MSIX: 26.721.4979.0 x64
• OS: Windows 10 Pro 10.0.19045
• GPU: NVIDIA GeForce RTX 3060, driver 32.0.15.6636
• Trigger: Browser Use navigated directly to https://dash.cloudflare.com/
• The browser tool reached DOMContentLoaded, returned the page title and a screenshot successfully, then the entire Codex app exited before the task could produce a final response.
• Four seconds after navigation, Microsoft-Windows-CodeIntegrity/Operational recorded Event ID 3033:

ChatGPT.exe attempted to load ...\app\vk_swiftshader.dll that did not meet the Microsoft signing level requirements.

• Get-AuthenticodeSignature reports the packaged vk_swiftshader.dll as NotSigned.
• No matching Application Error / WER event was recorded in that interval.

This shows the issue is not limited to Windows 11/25H2. It reproduces on Windows 10 build 19045 with the same Codex build and the same Code Integrity rejection. Using the Cloudflare REST API/CLI instead of the in-app browser avoids the crash. Raw session/log files are not attached because they contain private task data, but redacted timestamps and structural details are available if useful.

Additional reproducible evidence from Windows 11 / Codex Desktop 26.721.4979.0:

• Codex repeatedly exits during in-app browser activity, including from short/new tasks; it is not dependent on a long task history.
• Windows Code Integrity Operational logged 13 matching Event ID 3033 entries on July 27–28. Each says ChatGPT.exe attempted to load the bundled vk_swiftshader.dll, which did not meet Microsoft signing-level requirements.
• At the latest exact occurrence (2026-07-28 11:30:51 America/Panama), the Codex log immediately recorded GPU process exit code 101457950, followed by GPU relaunch failure exit code 18.
• I reviewed 20 GPU crashes; 19 immediately followed in-app-browser lifecycle activity. The latest occurred when an authenticated dashboard webview was opened and the app-wide in-app-browser tab count had reached 23.
• The package reports status Ok; storage, RAM, and GPU driver checks did not reveal a corresponding hardware failure. Driver update, restart, GPU-routing changes, and app repair did not resolve it.
• Windows system logs showed orderly RuntimeBroker-initiated shutdowns and no corresponding Kernel-Power 41 / unexpected-power-loss event; the recurring incident is the Codex process closing, not a hardware power loss.

This provides a consistent causal chain: in-app browser initializes the bundled SwiftShader/GPU path → Windows Code Integrity rejects vk_swiftshader.dll → GPU child exits and relaunch fails → the desktop app closes.

Please fix the Windows MSIX signing/catalog treatment of SwiftShader (or the Chromium GPU fallback policy), and ensure a GPU-child failure cannot terminate the entire application. Please also advise whether there is a supported temporary workaround that does not weaken Windows Code Integrity. Disabling Windows security or modifying the signed MSIX is not an acceptable workaround.

Impact: repeated loss of active work and multiple reinstall/repair attempts across paid use.

Additional reproduction on 2026-07-28:

• Codex Windows desktop package: 26.721.4979.0
• The app crashed three consecutive times during an integration workflow involving browser navigation/authentication (DeepL and Cloudflare).
• Each crash interrupted the active task and forced the workflow to be resumed manually in an external Chrome window.
• The repeated failures wasted substantial user time, effort, and paid AI credits.
• After the crashes, local shell operations also surfaced windows sandbox: CreateProcessWithLogonW failed: 5, matching this report.

Expected: an embedded-browser/GPU child-process failure should be isolated and recoverable; it must not terminate the whole Codex app or consume paid usage for work that cannot complete.

I can independently confirm the same issue on Windows with the exact package and failure sequence.

Environment

• Codex Microsoft Store package: OpenAI.Codex_26.721.4979.0_x64__2p2nqsd0c76g0
• Windows 11 25H2, build 26200.8894
• NVIDIA GeForce RTX 3060 Laptop GPU and Intel Iris Xe, both on recent drivers

Repeated failure sequence

This occurred five times on 2026-07-28:

1. The Codex in-app browser created or restored a browser webview.
2. Microsoft-Windows-CodeIntegrity/Operational recorded Event ID 3033 for packaged ChatGPT.exe attempting to load vk_swiftshader.dll.
3. The event data showed RequestedPolicy=8, ValidatedPolicy=1, and Status=0xC0000428.
4. At the same timestamp, the Codex desktop log recorded:

``text
Recoverable Chromium child process gone exitCode=101457950 processType=GPU reason=crashed serviceName=GPU
Recoverable Chromium child process gone exitCode=18 processType=GPU reason=launch-failed serviceName=GPU
``

1. Codex exited or became unusable.

Repairing or reinstalling the Microsoft Store package restores startup temporarily, but it stages the same package version and DLL, so the failure recurs when browser state is restored.

Additional observations

• The MSIX package reports Status: Ok and SignatureKind: Store.
• Get-AuthenticodeSignature reports the individual bundled vk_swiftshader.dll as NotSigned.
• No concurrent out-of-memory event, display-driver reset, standard Application Error/AppHang, or Crashpad dump was recorded in these incidents.
• The failure persisted after reboot and repeated Store repair/reinstallation.
• Avoiding the in-app browser prevents recurrence.

This currently makes Browser Use effectively unusable on the affected Windows installation and repeatedly interrupts active work. Code Integrity should remain enabled; the packaged GPU/SwiftShader initialization path needs to fail safely or use a binary accepted by the child-process signature policy.