Codex desktop: suspected false-positive safety checks repeatedly delay non-cybersecurity and non-biological work

I frequently encounter the additional-safety-check notice during ordinary work that involves neither cybersecurity nor biology. The repeated checks noticeably lengthen my working time. Please investigate whether these requests are being incorrectly routed into biological/cybersecurity safety review.

Observed behavior

• Product: Codex desktop app.
• Selected model: GPT-6 Astra; the UI displays the highest reasoning setting.
• UI language: Simplified Chinese.
• The notice appears repeatedly during work unrelated to cybersecurity or biology.
• Two screenshots of one affected task show the notice with total processing times of 6 minutes 29 seconds and 8 minutes 43 seconds.
• These are total task times; the duration attributable specifically to the safety check has not been measured.

Exact notice:

正在进一步审慎考虑此请求
如果不想等待，可以改用更快的模型重试。该模型处理复杂请求的能力可能较弱。

English translation: "This request is being considered further with extra care. If you do not want to wait, you can retry with a faster model. That model may be less capable of handling complex requests."

The UI offers Retry, Not interested, and Learn more. Learn more opens the OpenAI Help Center article Additional safety checks for biological and cybersecurity requests in ChatGPT, Codex, and the API.

Reproduction status

This is a recurring user-reported problem, supported by screenshots of an affected task. A deterministic minimal reproduction and an exact incidence rate are not yet available.

Expected behavior

Please investigate possible false positives affecting requests unrelated to cybersecurity or biology, correct misclassification if confirmed, and reduce avoidable waiting time.

Evidence limits

The screenshots establish the visible notice and total elapsed times. They do not establish the internal trigger or prove a backend defect. This report concerns suspected misclassification and recurring latency.

The report was prepared with assistance. Account details, private work descriptions, local paths, session logs, and original screenshots are omitted from this public draft.

Start by reproducing the notice in the Codex desktop app with work unrelated to cybersecurity or biology, recording the request context and total timing. Compare repeated affected requests with unaffected ones, then determine whether the behavior is a false-positive safety classification; done means the trigger is explained and any confirmed misclassification or avoidable delay is addressed.