In the past 24 hours, the Riemann hypothesis proof drew an authoritative response, the AI inference supply chain was hit by a large prompt leak, OpenAI issued new prompting guidance for GPT-6.2 developers, and Qwen4 Max arrived on Together AI. Here are the 5 stories that matter most.
Fields Medalist Maynard: The Riemann Proof Looks Right
On November 22, Fields Medalist and Oxford professor James Maynard said he believes the Riemann hypothesis proof completed by Anthropic mathematician Levent Alpöge with Claude is correct. On November 21, the Lean and Mathlib communities confirmed the proof passed formal verification. Yesterday's briefing covered how the proof was released. Maynard stressed that understanding why a theorem holds matters far more than proving it does.
OpenRouter Drops Three Providers; 125 Endpoints Lose Zero-Retention Status
On November 22, model-routing platform OpenRouter said that after security firm Dragos disclosed data leaks at inference providers, it has removed three providers, stripped 125 model endpoints of Zero Data Retention eligibility, and notified customers who may be affected. OpenRouter stressed its own platform was not breached and that some endpoints were removed as a precaution.
Together AI Reports a Prompt Leak: A Coding Agent Committed a Repository Credential
On November 21, Together AI reported a security incident: while developing the public xorl repository, an AI coding agent committed a credential for a private Hugging Face repository, and an unauthorized party used it to download prompt samples submitted in June and July 2026 to MiniMax M3, GLM-5.2, Kimi K2.6 and other models, then searched them for other services' credentials. The company is cooperating with an external forensics firm and law enforcement.
OpenAI Publishes GPT-6.2 Prompting Guidance; Editing History Drops Reasoning
On November 21, OpenAI published new prompting guidance for GPT-6.2 Astra and Sol in its developer documentation: Astra answers more concisely and directly by default, and when intent is unclear it should push ahead rather than wait; code output is leaner by default, with new tests added only when needed. The guidance warns that editing conversation history discards earlier reasoning content, hurting answer quality and raising token use.
Qwen4 Max Lands on Together AI, Fixing a Whole Repo's Bug in One Call
On November 21, Together AI said Alibaba's flagship model Qwen4 Max is live on its serverless inference platform as Qwen/Qwen4-Max, with a 1M-token context window and 30% off through November 22. In a test, the model received a bug report plus all 344 files of the Tornado framework — 650,000 tokens — in a single call and returned a patch that passed the project's tests within 72 seconds, for about $0.69. Qwen4 Max was released on November 16.