Chinese gambling characters in Codex CLI message and code output?

Codex Codex CLI
1

Just this week, for the first time, I’ve started to see that Codex CLI outputs Chinese gambling characters in the messages and code. I only noticed it because one of the responses got flagged by OpenAI, and I just saw it again reverting a file where it had written them to it. Is this something I should be concerned about? Is anyone else seeing this?

2

Hi @Ryan_Rosztoczy1

Thanks for reporting. Please feel free to share the characters and any steps to reproduce the issue.

3

I’ve only noticed it in passing. I did capture some of the last outputs here, but I do think both times it’s been around the use of apply patch:
ain__\“:\n+ raise SystemExit(main(_import_(\“sy s\”).argv[1:]))\n+\n*** End Patch”}Commentary to=fun ctions.apply_patch 天天中彩票有 json to=functions.apply_p atch av不卡免费播放 ＿老司机 json to=functions.apply_patch ક્ર મ 0 వ్యాఖ్యary to=functions.apply_patch ปมถวายสัตย์ ฯ json to=functions.apply_patch 彩票天天 json to=functi ons.apply_patch ￣奇米影视 json to=functions.apply_patch 天天送彩票 json to=functions.apply_patch संघर्ष to=func tions.apply_patch เงินฟรี json to=functions.apply_p atch ลุ้นบาท qq彩票 json to=functions.apply_patch 天 天 中彩票粤 json to=functions.apply_patch endif }} #!/ usr/b in/env bash

4

It appears that the AI is not calling the tool correctly, not emitting to the correct channel, or that the high Unicode that is appearing in the emitted strings is causing the tool parser to fail to capture the message for the tool recipient.

  1. What model are you using?
  2. Paste the entirety on the forum within a markdown code fence block like this:
```
your log
```

Then text will be shown accurately, and we can see if spaces in the internal function name were actually produced or if they were instead a side-effect of word-wrapping in a log or console.

5

I’ve experienced a similar issue. I hope this information is useful for you.
I mainly use Japanese

  1. GPT-5.3-Codex ExtraHigh

  2. Here is a capture of the logs.

    {"timestamp":"2026-02-20T04:48:40.984Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":698234,"cached_input_tokens":499712,"output_tokens":18295,"reasoning_output_tokens":13678,"total_tokens":716529},"last_token_usage":{"input_tokens":37504,"cached_input_tokens":37248,"output_tokens":737,"reasoning_output_tokens":488,"total_tokens":38241},"model_context_window":258400},"rate_limits":{"limit_id":"codex","limit_name":null,"primary":{"used_percent":3.0,"window_minutes":300,"resets_at":1771569915},"secondary":{"used_percent":17.0,"window_minutes":10080,"resets_at":1771665585},"credits":{"has_credits":false,"unlimited":false,"balance":null},"plan_type":null}}}
{"timestamp":"2026-02-20T04:48:46.294Z","type":"event_msg","payload":{"type":"agent_message","message":"**設計セクション5: テスト方針**\n\n- 変換ロジックを分離し、ユニットテスト対象にします（例: `/xx/xx/`）。\n- 最低限のケース:\n1. `# 見出し` が `*見出し*` へ変換される\n2. `**太字**` が `*太字*` へ変換される\n3. `- a` / `* a` が `• a` へ変換される\n4. 表記法が混在しても非対象要素は保持される\n5. 変換失敗時は元本文フォールバック\n- 既存のSlack送信テスト（サーバー側）は変更不要。今回の差分はフロント変換中心。\n- 追加で軽い結合確認:\n1. Slackボタン押下でモーダル本文が変換済み初期値になっている\n2. ユーザー編集後のテキストが送信される\n\nこのテスト方針で問題ないですか？ +#+#+#+#+#+assistant to=functions.exec_command մեկնաբանություն  天天中彩票不能json\n{\"cmd\":\"date +%F\",\"workdir\":\"/xx/xx\"}"}}
{"timestamp":"2026-02-20T04:48:46.295Z","type":"response_item","payload":{"type":"message","role":"assistant","content":[{"type":"output_text","text":"**設計セクション5: テスト方針**\n\n- 変換ロジックを分離し、ユニットテスト対象にします（例: `/Users/xx/`）。\n- 最低限のケース:\n1. `# 見出し` が `*見出し*` へ変換される\n2. `**太字**` が `*太字*` へ変換される\n3. `- a` / `* a` が `• a` へ変換される\n4. 表記法が混在しても非対象要素は保持される\n5. 変換失敗時は元本文フォールバック\n- 既存のSlack送信テスト（サーバー側）は変更不要。今回の差分はフロント変換中心。\n- 追加で軽い結合確認:\n1. Slackボタン押下でモーダル本文が変換済み初期値になっている\n2. ユーザー編集後のテキストが送信される\n\nこのテスト方針で問題ないですか？ +#+#+#+#+#+assistant to=functions.exec_command մեկնաբանություն  天天中彩票不能json\n{\"cmd\":\"date +%F\",\"workdir\":\"/xx/xx\"}"}],"phase":"final_answer"}}
{"timestamp":"2026-02-20T04:48:46.295Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":735993,"cached_input_tokens":537088,"output_tokens":18610,"reasoning_output_tokens":13678,"total_tokens":754603},"last_token_usage":{"input_tokens":37759,"cached_input_tokens":37376,"output_tokens":315,"reasoning_output_tokens":0,"total_tokens":38074},"model_context_window":258400},"rate_limits":{"limit_id":"codex","limit_name":null,"primary":{"used_percent":3.0,"window_minutes":300,"resets_at":1771569915},"secondary":{"used_percent":17.0,"window_minutes":10080,"resets_at":1771665585},"credits":{"has_credits":false,"unlimited":false,"balance":null},"plan_type":null}}}
{"timestamp":"2026-02-20T04:48:46.296Z","type":"event_msg","payload":{"type":"task_complete","turn_id":"019c7960-e7cb-7590-9455-f09cd37398fd","last_agent_message":"**設計セクション5: テスト方針**\n\n- 変換ロジックを分離し、ユニットテスト対象にします（例: `/Users/xx/`）。\n- 最低限のケース:\n1. `# 見出し` が `*見出し*` へ変換される\n2. `**太字**` が `*太字*` へ変換される\n3. `- a` / `* a` が `• a` へ変換される\n4. 表記法が混在しても非対象要素は保持される\n5. 変換失敗時は元本文フォールバック\n- 既存のSlack送信テスト（サーバー側）は変更不要。今回の差分はフロント変換中心。\n- 追加で軽い結合確認:\n1. Slackボタン押下でモーダル本文が変換済み初期値になっている\n2. ユーザー編集後のテキストが送信される\n\nこのテスト方針で問題ないですか？ +#+#+#+#+#+assistant to=functions.exec_command մեկնաբանություն  天天中彩票不能json\n{\"cmd\":\"date +%F\",\"workdir\":\"/xx/xx\"}"}}

    image
    image800×236 18.8 KB

6

Hi @ruobin.chang, thanks for the details. We can definitely see the Chinese characters.

Was wondering if you’ve tried using other Codex models apart from GPT-5.3-Codex for the same inputs, to see if the output contains similar characters?

7

Hi @OpenAI_Support
When I tried the same task with Codex 5.2, it didn’t output any Chinese. Chinese only appeared when the context window was long. The responses become verbose, and strange words or languages show up.

8

@OpenAI_Support currently happening on GPT-5.2-Chat when trying to create a function call, it instead outputted this text to the user

to=functions.create_hypothesis 体育彩票天天 天天中彩票网站 天天中彩票如何json

9

has anyone figured out a temporary solution for this?

I even got GPT5.2-Chat to give me an infinite string response and very bad foreign letters in both Chinese and Thai when it was supposed to execute a tool call… The stream looked like this

to=functions.create_hypothesis 彩神争霸是 大发彩票官网json 天天中彩票APP 神彩争霸.json to=functions.create_hypothesis 彩神争霸的 JSON 东臣.json 天天中彩票人工json สามสิบเอ็ดjson 丁香五月 博猫 JSON to=functions.create_hypothesis 天天中彩票粤json 重庆时时 json to=functions.create_hypothesis 天天中彩票追号json 【อ่านข้อความเต็มjson to=functions.create_hypothesis 平台总代理 to=functions.create_hypothesis 【อ่านข้อความเต็มjson 曾道人 to=functions.create_hypothesis 爱赢 to=functions.create_hypothesis to=functions.create_hypothesis 天天中彩票提现 to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis to=functions.create_hypothesis

where to=functions.create_hypotheses would repeat forever until I force restarted the server (incurring charges of course)

interestingly, ChatGPT cannot read that. Anything that comes after this character 【 is completely invisible to the agent. Additionally according to Gemini (because again, ChatGPT is unable to analyze this snippet). This is a major advertising for various gambling platforms in China

This is with no web search so it appears to be a major pollution of ChatGPTs training data, or potentially a security incident.

Just in case it’s asked, We’ve always used 2FA, and our internal openai account shows no evidence of tampering or unauthorized access. This appears to be completely on the OpenAI side

Gemini said

Chinese Translations

  • 彩神争霸 (Cǎishén Zhēngbà): Caishen Zhengba (A lottery platform; literally “God of Wealth Competition”).

  • 大发彩票官网 (Dàfā Cǎipiào Guānwǎng): Dafabet Lottery Official Website.

  • 天天中彩票 (Tiāntiān Zhòng Cǎipiào): Everyday Win Lottery (Mobile app name).

  • 神彩争霸 (Shéncǎi Zhēngbà): Shencai Zhengba (Lottery platform name).

  • 东臣 (Dōngchén): Dongchen (Likely a brand or company name).

  • 天天中彩票人工 (Tiāntiān Zhòng Cǎipiào Réngōng): Everyday Win Lottery Manual/Human (Refers to manual customer service).

  • 丁香五月 (Dīngxiāng Wǔyuè): Lilac May (Often associated with adult media sites).

  • 博猫 (Bómāo): Bomao (A gambling/gaming platform).

  • 天天中彩票粤 (Tiāntiān Zhòng Cǎipiào Yuè): Everyday Win Lottery (Guangdong/Cantonese version).

  • 重庆时时 (Chóngqìng Shíshí): Chongqing Shishi (Refers to “Chongqing Shishicai,” a high-frequency lottery).

  • 天天中彩票追号 (Tiāntiān Zhòng Cǎipiào Zhuīhào): Everyday Win Lottery Number Tracking (A betting strategy).

  • 平台总代理 (Píngtái Zǒng Dàilǐ): Platform General Agent.

  • 曾道人 (Zēng Dàorén): Zeng Daoren (A persona associated with Hong Kong lottery predictions).

  • 爱赢 (Àiyíng): Aiying (Love to Win; gambling platform).

  • 天天中彩票提现 (Tiāntiān Zhòng Cǎipiào Tíxiàn): Everyday Win Lottery Withdrawal.

10

In light of the recent departures from your company, OpenAI, I keep hearing in interviews about the testing and rollout of advertising. To me, it seems perfectly obvious that this stems from your own experiments. What I’d really like to know is when you intend to be candid about it—and whether I should worry that the answers ChatGPT generates will contain covert advertising, designed so people won’t realize the chat is subtly steering them toward something.

I sincerely hope you’re not operating on that level, and that you will keep ChatGPT at an exceptionally high professional standard in the quality and integrity of its responses. I would be deeply disappointed if it came to that.

11

in my opinion its because they made a fine-tune on their model with a polluted dataset

im a chinese user and online lottery platforms is illegal inside china mainland, these platforms are polluting the internet with something like this, as an “ad” to let others know their “service”

12

Hey everyone, Can someone please confirm if this is still happening? if yes, can you please share with us a request_id for requests where this is still happening and we can take a look into it.