Agent-initiated Browser Use can trigger Code Integrity 3033, GPU relaunch failure, and Store remediation

What version of the Codex App are you using (From “About Codex” dialog)?

26.721.4979.0 (OpenAI.Codex_26.721.4979.0_x64__2p2nqsd0c76g0)

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

On two occasions, an active Codex task automatically activated and interacted with an already-existing Browser Use tab. I did not manually open a conversation or webpage at the time of either failure. Immediately afterward, Windows Code Integrity rejected the packaged Chromium SwiftShader library, the GPU child process crashed, its replacement failed to launch, and the packaged app entered a remediation state. Codex could not be reopened normally until Microsoft Store repaired/restaged the package.

The same sequence occurred on 28 and 29 July 2026 on build 26.721.4979.0:

Codex Browser Use activates/reuses an existing tab
  -> Code Integrity Event 3033 for packaged app\vk_swiftshader.dll
  -> GPU child exits: exitCode=101457950, reason=crashed
  -> replacement GPU child exits: exitCode=18, reason=launch-failed
  -> packaged app container is destroyed / package requires remediation
  -> Microsoft Store repairs or restages the same package version
  -> package status returns from 2 to 0 and Codex launches again

The relevant Code Integrity fields were identical in both incidents:

Event ID: 3033
ImageName: ...\OpenAI.Codex_26.721.4979.0_x64__2p2nqsd0c76g0\app\ChatGPT.exe
FileName:  ...\OpenAI.Codex_26.721.4979.0_x64__2p2nqsd0c76g0\app\vk_swiftshader.dll
RequestedPolicy: 8
ValidatedPolicy: 1
Status: 3221226536 (0xC0000428)

This does not appear to be local DLL tampering. The outer AppxSignature.p7x validates successfully, and all 83 uncompressed 64-KiB blocks of the current app\vk_swiftshader.dll match the hashes declared in AppxBlockMap.xml. The current file SHA-256 is:

2C8F89FA26945207260B9B019F7E1A2F4F6820741B783561F729DC13C7E82C7E

The Store recovery was a same-version remediation, not an upgrade to a newer Codex build. In both cases Microsoft Store found an applicable operation, downloaded/restaged 26.721.4979.0, and Windows later recorded package status changing from 2 to 0.

What steps can reproduce the bug?

The following is the observed reproduction sequence. It produced the complete failure twice, but is not yet a deterministic minimal reproducer.

1. Run Codex Desktop 26.721.4979.0 on Windows while signed in.
2. Have an existing Browser Use tab retained from an earlier task. In my two incidents this was an authenticated third-party web workflow; the URL and account identifiers are redacted.
3. Start or continue a Codex task that requires the agent to inspect and interact with that existing tab. Do not manually open a new conversation or webpage.
4. Let Codex activate the retained tab and attach its Browser Use debugger. The retained session shows browser tab inspection/DOM snapshot operations and click or evaluate operations immediately before the failures.
5. Observe Windows Code Integrity Event 3033 for the packaged vk_swiftshader.dll.
6. Observe these consecutive Codex log entries:

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

1. Observe the Codex window disappear and normal relaunch fail or return to an unusable package state.
2. Open Microsoft Store and repair/update Codex. The Store may download the same version rather than a newer version.
3. After the package state returns to normal, launch Codex successfully.

Observed timing:

| Local time (UTC+08:00) | Evidence |
|---|---|
| 2026-07-28 21:33:55 | Existing Browser Use tab activated; debugger listener registered |
| 2026-07-28 21:34:16 | Code Integrity 3033, record ID 1614080 |
| 2026-07-28 21:34:17 | GPU crash followed by GPU launch failure |
| 2026-07-28 21:38:33 | AppX registration repair begins |
| 2026-07-28 21:40:47 | Same-version Store remediation completes; status 2 -> 0 |
| 2026-07-29 10:43:16 | Existing Browser Use tab activated; debugger listener registered |
| 2026-07-29 10:43:17 | Code Integrity 3033, record ID 1614083 |
| 2026-07-29 10:43:18 | GPU crash followed by GPU launch failure |
| 2026-07-29 10:43:29 | AppX registration repair begins |
| 2026-07-29 11:03:27 | Same-version Store remediation completes; status 2 -> 0 |

Reproduction limits and negative controls:

• The complete chain was observed twice on the current build.
• A later Browser Use session produced additional 3033 events for vk_swiftshader.dll and vulkan-1.dll without crashing the app. Therefore Event 3033 alone is not sufficient to reproduce the full failure.
• A controlled test that force-terminated all 12 package-owned ChatGPT.exe processes during an active task did not reproduce the problem; the app relaunched on the first attempt. Generic abnormal termination is therefore not the demonstrated trigger.
• I have not forced SwiftShader using Chromium command-line flags because doing so against the primary packaged app may deliberately put it back into the unrecoverable/remediation state. No unexecuted command-line experiment is presented here as a confirmed PoC.

Session ID: redacted because the retained task contains authenticated third-party workflow details. Token-limit and context-window usage do not appear relevant to the native GPU/package failure.

What is the expected behavior?

Browser Use should not cause a packaged Chromium component to be rejected by the GPU process's Code Integrity policy. If the GPU process crashes, Codex should recover with a new GPU process or fall back safely. It should not place the AppX package into a state that requires Microsoft Store remediation before the app can launch again.

Additional information

• Current GPU process command line contains --type=gpu-process but no explicit --use-angle or SwiftShader flag. The SwiftShader load therefore appears to be selected internally rather than requested by the user.
• The GPU process has BlockRemoteImageLoads=ON and BlockLowLabelImageLoads=ON. The rejected DLL is inside the installed Store package, not a remote or low-integrity path.
• No correlated display-driver reset, WHEA, NVIDIA, AMD, Application Error 1000, or WER crash event was found for either incident. The direct evidence is Code Integrity 3033, the Codex GPU-child logs, and the AppX/Store remediation timeline.
• Graphics environment: NVIDIA GeForce RTX 3050 Laptop GPU (31.0.15.2656), AMD Radeon Graphics (31.0.12044.3), and Oray virtual display driver (17.1.58.818). These drivers are old, but no driver-reset event correlates with either failure, so I cannot attribute the bug to a driver.
• The affected authenticated URL, OAuth state/nonce, identity-verification token, local username, device name, user SID, and submission identifiers must remain redacted from public logs.
• This is reported as a Codex Windows reliability/package-integration bug. I am not claiming a security boundary bypass or asking Codex to bypass Windows Code Integrity.