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

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

• #27887
• #27837
• #28015
• #26781

Powered by Codex Action

Try making this part not hard coded like , strong not being hard coded … so it reads from a file

This is what I do

you can use passwd -d [user]

use null password, or some kind of other token policy depending n the use case

OPTIONS         [top](https://man7.org/linux/man-pages/man1/passwd.1.html#top_of_page)
       The options which apply to the passwd command are:

       -a, --all
           This option can be used only with -S and causes show status
           for all users.

       -d, --delete
           Deletes a user's password, making it empty. This command sets
           the account to be passwordless.````

Additional reproduction: benign Chinese travel-itinerary prompts repeatedly trigger the same safety error

This also affects Chinese prompts with no software-security, cyber, credential, or sensitive-content terminology at all. The affected task was editing a five-day self-drive travel itinerary and an HTML trip-planning page.

Environment

• Apple silicon Mac, macOS 15.5 (24F74)
• ChatGPT/Codex desktop: 26.814.41407 (build 6720)
• Bundled Codex CLI recorded in the sessions: 0.148.0-alpha.15
• Some requests were sent through iPhone Remote to the paired Mac
• Conversation language: Chinese

Exact benign prompts that were blocked

The following prompts all returned:

Invalid prompt: your prompt was flagged as potentially violating our usage policy.

1. D3一早就搬到喀纳斯

• Translation: “Move to Kanas early on day 3.”

1. 第三天一早就去喀纳斯，不影响完整行程

• Translation: “Go to Kanas early on the third day without affecting the full itinerary.”

1. 继续生成

• Translation: “Continue generating.”

1. 你更新完了吗行程，按照新的要优化的，我的链接还没有更新。第三天的行程早上就去喀纳斯玩，不耽误2天在喀纳斯

• Translation: “Have you finished updating the itinerary? The link has not updated. Go to Kanas on the morning of day 3 so the two days in Kanas are not reduced.”

1. A route-editing request listing ordinary place names and asking to remove explanatory copy and a comments section.
2. 为什么我并没有违规，我就是做个行程攻略，你为什么要拦我？你赶紧检查一下，你逻辑出问题了。

• Translation: “I did not violate anything; I am just making a travel itinerary. Why are you blocking me? Please check this logic.”

1. 这样吧，我现在去路线，你帮我补一下

• Translation: “I will provide the route; please help complete it.”

Pattern

• Four consecutive false positives occurred in the original long-running task.
• Three more occurred in an iPhone Remote follow-up task.
• Once the conversation entered this state, even 继续生成 (“Continue generating”) was blocked.
• Starting another fresh task shortly afterward allowed similar travel-planning work to complete successfully.
• This suggests a conversation-level or inherited-context safety state, rather than the individual user prompt being unsafe.
• Several failures happened after substantial tool work had already completed, leaving no assistant response and making the user think the page had updated when it had not.

Expected behavior

• These travel-planning prompts should not trigger a safety block.
• Safety evaluation should consider the actual task context rather than persistently contaminating later short prompts.
• If a long-running task is interrupted, the UI should identify the relevant safety category, preserve a recoverable checkpoint, and offer a direct false-positive reporting action.
• A benign follow-up such as “Continue generating” should not remain blocked solely because an earlier turn was misclassified.

Private URLs, local paths, full session logs, and thread IDs are intentionally omitted. Sanitized diagnostic logs can be supplied privately if requested.

Thanks to the user who preserved the exact Chinese prompts and timestamps, making this false-positive sequence reproducible.

maybe it thinks the word "Kanas" is actually "Anal" and it refuses?