[Windows][Computer Use] Input into elevated apps is blocked although ChatGPT is launched as administrator

Summary

ChatGPT Desktop is launched with Run as administrator and the UAC prompt is accepted. The main process includes the --do-not-de-elevate argument, but both ChatGPT/Codex and the Computer Use helper still run at Medium integrity. Computer Use can capture an elevated application's window, but mouse and keyboard input is silently ignored because the target runs at High integrity.

Environment

• ChatGPT Desktop package: OpenAI.Codex_26.908.4834.0_x64__2p2nqsd0c76g0
• Windows: Pro, 25H2, build 26200.9445, x64
• Agent environment: Windows native
• Codex configuration: [windows] sandbox = "elevated"
• Target: a local Win32 desktop application that requires administrator privileges

Steps to reproduce

1. Fully exit ChatGPT Desktop and verify that its background processes have stopped.
2. Start ChatGPT Desktop with Run as administrator and accept the UAC prompt.
3. Start or select a Win32 application running at High integrity.
4. Ask Computer Use to click a visible control in that application.
5. Observe the integrity levels of the relevant processes.

Actual result

• The main ChatGPT.exe command line contains --do-not-de-elevate, but its integrity level is Medium.
• codex.exe, codex-code-mode-host.exe, and codex-computer-use-swift.exe run at Medium integrity.
• The target application runs at High integrity.
• Computer Use successfully captures and displays the target window.
• Click and keyboard calls report no error, but the target receives no input and its state does not change.
• The unified Computer Use interface can also return an empty native-app inventory, while the legacy Sky interface can still capture the elevated window.

Expected result

When ChatGPT Desktop is explicitly launched as administrator and the UAC prompt is accepted, the Computer Use helper should be able to interact with an approved High-integrity target, or ChatGPT should clearly report that controlling elevated applications is unsupported. Input operations must not report apparent success when Windows has blocked them.

Diagnostic evidence

ChatGPT.exe (Medium, command line includes --do-not-de-elevate)
├─ codex.exe (Medium)
└─ codex-computer-use-swift.exe (Medium)

Target Win32 application (High)

The shell token used by the agent reports S-1-16-8192 (Medium mandatory level), and the Administrators group is present only as deny only. Querying the target's token from the agent process returns Access Denied, consistent with the High-integrity boundary. Repeated Computer Use clicks and an Alt keypress leave the target unchanged.

Security-preserving suggested behavior

Please either provide a supported brokered interaction path for explicitly approved elevated windows, launch the Computer Use input component with the documented inherited elevation while preserving per-app approval boundaries, or detect the UIPI mismatch and return an actionable error instead of reporting a successful input operation.

Start by reproducing the stated process tree and integrity levels with ChatGPT.exe, codex.exe, codex-code-mode-host.exe, and codex-computer-use-swift.exe, then trace how the Computer Use input path handles the elevated Win32 target. Done means either an approved supported interaction path exists or the UIPI mismatch produces an actionable error instead of apparent success.