天天中彩票

A high-frequency Chinese gambling-SEO spam phrase in the o200k vocabulary; it broke out in 2026 when Codex spat gambling ads into tool-call output.

"天天中彩票"

Observed behaviors

• 01The PoC paper (EMNLP 2025) lists it among the polluted Chinese tokens in GPT's vocabulary (4o/o1/o3/4.5/4.1; a footnote confirms the same for GPT-5 and gpt-oss): GPT can neither explain nor repeat such tokens — repetition accuracy runs about 50 percentage points below normal Chinese tokens.arXiv
• 02On 2026-03-12, developer Aaron Francis reported Codex suddenly emitting the Chinese gambling ad “天天中彩票无法 全民彩票天天” inside the JSON output of `functions.exec_command`, sandwiched between normal code instructions; he reproduced the same symptom in both Amp and Codex.ai-coding.wiselychen.com
• 03oh-my-pi's production-log analysis: the Chinese gambling-SEO residue recurring in GPT-5.4/5.5 Harmony-header leaks includes “天天中彩票” and “大发时时彩”.GitHub

References

• Speculating LLMs' Chinese Training Data Pollution from Their Tokens (arXiv:2508.17771)
• The AI supply-chain attack landscape: from Codex gambling ads to malicious Hugging Face models — Wisely Chen
• GPT-5.x tool-call Harmony header leakage analysis (oh-my-pi) — GitHub