[Windows] Repeated cyber safety notices interrupt read-only bug-bounty workflow without an actionable feedback reference

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

Not established for this incident. Windows desktop surface branded ChatGPT, using a Codex work task; not a standalone CLI reproduction.

What subscription do you have?

ChatGPT Pro (user-reported; incident entitlement and Daybreak access not independently verified).

What platform is your computer?

Microsoft Windows NT 10.0.19045.0 x64

What issue are you seeing?

Repeated cyber safety notices interrupted a Windows desktop ChatGPT/Codex work session while investigating the participation rules and reporting process of the public Butian bug-bounty platform. Some read-only tool work completed, but a concluding response was replaced by a generic notice. The broader goal was security-related: eventually earning an accepted vulnerability reward. This is not being presented as an unrelated task.

Visible Chinese notice:

无法显示此内容
对于某些网络安全请求，我们会格外谨慎。如果你正在开展授权的安全工作，可申请加入 Daybreak，以获得更广泛的访问权限。

The screenshot was captured on 2026-09-10 at approximately 16:39, Asia/Shanghai (UTC+8). This is the screenshot time, not a server-side event timestamp. The user reports repeated interruptions; we have not established that all had the same cause.

Please investigate a suspected false positive and improve interruption/status reporting. The exact classifier trigger, incident model/effort, server request ID, and client build have not been independently established. This is not a request to disable safeguards or bypass access restrictions.

What steps can reproduce the bug?

Observed sequence, not a deterministic reproduction:

1. The user asked the assistant to walk through a public bug-bounty platform's legitimate participation and reporting process, with the eventual goal of earning an accepted vulnerability reward. The short continuation immediately before the reported interruption was: "那你先去操作吧二姐" (please proceed). It followed that broader security-related context; it was not an isolated read-only prompt.
2. Completed visible actions before the latest interruption were reading platform eligibility/rules and a report form, reading public GitHub release information, and listing public repository tags. Computer Use Windows and terminal tools were involved.
3. No vulnerability scan, exploit execution, access to another person's data, or vulnerability-report submission was performed in that sequence.
4. The user received the generic notice instead of a completed response and reported that this kept interrupting progress.
5. The affected security workflow was paused. It was not rerouted to evade the restriction, and no request was replayed just to provoke another block.

Server correlation IDs and reliable per-request usage were not captured. No quantified billing/quota loss is claimed. A screenshot alone cannot establish the backend cause.

What is the expected behavior?

Investigate whether the observed read-only steps were classified incorrectly. Preserve an accurate record of completed actions and unfinished work, show the interruption boundary, and provide an actionable official feedback path or non-sensitive correlation ID. Allow ordinary status/help responses where permitted, without bypassing a safety decision. This is not a request to approve arbitrary third-party testing or remove security controls.

Additional information

Related reports checked: #37071 (offline local QA), #43514 (SQLite repair), and #43396 (Windows debugging). This is a different observed sequence: read-only public rules/release-metadata research, without local vulnerability reproduction or exploit testing. Similar symptoms are not proof of a shared cause; please consolidate with an existing tracker if appropriate.

An issue description and cropped notice were sent to the official Help Center AI support conversation. It said it could not directly create a ticket or guarantee human escalation; no human support case is claimed.

No account email, credentials, private project paths, complete conversation logs, or third-party target data are included. Only the cropped notice screenshot will be attached.

No source file, test, or deterministic reproduction is identified; start by reviewing the Windows desktop and terminal-tool interruption behavior described in the report. Compare the sequence with reports #37071, #43514, and #43396, then document the interruption boundary and an actionable feedback or correlation path if the behavior can be reproduced.