Summary

While using the ChatGPT Android app today (August 14, 2026), an unexpected fragment of Chinese text associated with online gambling/spam appeared inside the visible reasoning-summary UI (“Réfléchi pendant…” / “Thought for…”), attached to an otherwise standard internal execution step.

The string that appeared was:
申博太阳城 (Shenbo Sun City — a well-known online gambling/casino brand)

Attached directly to the normal step label:
Vérification du fichier d’état actuel 申博太阳城

Context and Observations

  • Environment: ChatGPT Android app.
  • Conversation topic: A professional compliance and audit project completely unrelated to China, gambling, casinos, lotteries, or promotional spam.
  • Location: The string appeared strictly within the reasoning-summary UI dropdown, not in the user prompt, tool arguments, or final assistant markdown response.
  • Codebase / Repository check: A search across all local project repositories, files, and referenced manifests yielded zero occurrences of this string.
  • Shareable reasoning trace: The corresponding shareable reasoning trace did not display this string.
  • Impact: No malicious execution or data loss was observed; the model proceeded with the task normally.
  • Screenshot: Attached below (client-sensitive URLs redacted, leaving the full reasoning UI and circled Chinese string intact).

Relation to Previous Reports

This behavior appears closely related to issues reported in the community over recent months (including the July 2026 reports regarding random Chinese lottery/gambling text like “天天彩票”, “天天中彩票”, and platform names appearing after heartbeat tokens or during tool calls in Codex/API workflows).

However, this specific case is notable because the fragment surfaced directly in the reasoning-summary UI on Android.

Questions

  1. Could this be an instance of the known multilingual generation / token leak issue previously acknowledged by the team, or could this point to some form of context boundary issue, cache pollution, or prompt injection?
  2. Is the reasoning-summary parser in the mobile client known to pick up extraneous model tokens under certain compaction or streaming conditions?

From a user perspective, having unsolicited gambling/spam strings surface inside internal reasoning steps is concerning, so sharing this report and visual evidence to help the OpenAI team investigate and trace the behavior.