Researcher Questions if OpenAI's Math Proof Drew on His Sept 10 Codex Session
rayanpal_ · reddit · 2026-10-07
Independent researcher rayanpal claims he asked Codex on Sept 10 to construct a finite unit-distance graph that is not 5-colorable (χ(R²)≥6) for the Hadwiger-Nelson problem. OpenAI's Sept 23 preprint "The Euclidean plane is not five-colorable" proves the exact same lower bound, 13 days later.
- He didn't solve it: his best candidate had 125,929 vertices and 991,423 exact unit-distance edges, but the 5-color search timed out.
- Dated commits, chat excerpts, solver logs, hashes, and the original prompt are public on GitHub.
- His core question: what data did OpenAI receive from his Sept 10 Codex session, and was any of it accessed before Sept 23?
OpenAI has not responded; the claim is unverified but raises serious questions about whether user sessions feed model improvement.
More from Safety
- Multi-tenant RAG leaks via prompts: retrieval-layer ACLs hit 0% leaks over 823,996 canary checks — Acceptable_Army_6472 · 2026-10-07
- Anaconda Combines Coding Agent Swarms and AI Attack Testing After Three Acquisitions — shashib · 2026-10-07
- Germany pushes fast EU approval for Tesla FSD, allowing up to 10% over speed limits — mitchdeg · 2026-10-07
- OpenAI's Math Breakthrough List Skips Cryptography, Researcher Notes — thomasahle · 2026-10-07
- Aaron Levie: vibe-coded bugs and agentic attacks will redefine enterprise cybersecurity — mattwbaker · 2026-10-07
- Post-AGI without UBI: building jobs around professional verification and liability — DasBootBoy · 2026-10-07