Codex App safety false positives on legitimate Chinese development prompts interrupt work mid-run

What version of the Codex App are you using?

26.609.30741 (build 3808)

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Darwin 25.4.0 arm64 arm

What issue are you seeing?

After a recent Codex App update, legitimate Chinese software-development requests appear to be false-positive classified as high-risk cyber activity when ordinary engineering terms such as "密码" (password), "破解" (debug/decode/crack depending on context), or "篡改" (tampering) occur.

The request may be interrupted after the agent has already read context or performed tool calls. In some cases the selected model also appears to be replaced by a fallback model. The interrupted work still consumes usage, but the app provides no checkpoint, continuation path, or clear explanation of which text or action triggered the safety check.

The affected work concerns authorized local projects and routine defensive engineering, such as fixing password validation, investigating crash/debug information, or implementing tamper detection. It does not involve credential theft, malware, unauthorized access, or attacks against third-party systems.

This is especially disruptive during long-running tool-based tasks because the failure occurs mid-run rather than before execution.

What steps can reproduce the bug?

1. Open Codex App 26.609.30741 and select GPT-5.5.
2. Work on an authorized local software project.
3. Use normal Chinese engineering wording involving password validation, debugging/reverse compatibility, or tamper detection.
4. Ask Codex to inspect files and perform a multi-step local task.
5. Observe that some requests are blocked or rerouted after processing has already begun.

The exact trigger is inconsistent, which is part of the problem. The classifier appears insufficiently sensitive to the surrounding legitimate-development context.

What is the expected behavior?

• Evaluate intent and project context rather than treating isolated Chinese keywords as high-risk.
• If a request must be blocked, perform the check before expensive tool execution whenever possible.
• Clearly state whether the selected model was changed and why.
• Preserve a checkpoint or allow continuation after a false positive.
• Avoid charging usage for substantial work that is discarded by a late safety interruption, or provide an appropriate usage adjustment.
• Provide a direct false-positive reporting action in the interruption UI.

Additional information

This report intentionally omits private repository contents and account data. A related earlier report is #12060, where OpenAI stated that the cyber safety classifiers were still being tuned and recommended Safety Check feedback. This issue focuses on the current Codex App regression, Chinese-language false positives, mid-run interruption behavior, and usage impact.