Lean community backs a machine-checked Riemann Hypothesis proof

On Friday 20 November, Anthropic mathematician Levent Alpöge said he'd found a proof of the Riemann Hypothesis "working with Claude". He posted it as a Lean codebase of about 17 million lines, and says it also covers the Generalized Riemann Hypothesis. The Lean and Mathlib community has since put out a statement saying it has "full confidence" the proof is correct. Dozens of people compiled the code on their own machines and got the same result. MIT's Andrew Sutherland, who keeps a list of those independent checks, called the result "a human triumph". He also said there's "much work to be done" before mathematicians understand how the argument works.

The Lean kernel confirms that the formal argument is valid. Mathematicians still have to check that the Lean statement matches the conjecture and to read the human write-up once it's published.

For builders: This is the largest public example yet of an AI system and a human producing a result that a proof checker can verify. Teams building reasoning or code agents should take note: pairing a model with a strict verifier like a type checker or test suite is what made this claim credible so fast.

Dragos catches inference providers leaking prompts and credentials

On 19 November, Dragos chief research officer HD Moore published the results of a five-month experiment. From July to November, the team sent roughly two million prompts to popular AI services, each seeded with unique AWS keys, hostnames and URLs, then watched which ones got used. Attackers had stolen data from inside at least three hosting providers. In one late-October breach, a couple of hundred planted AWS keys were used for tens of thousands of API calls. Resellers on one inference marketplace could read customer prompts through ordinary request logging. A Chinese AI lab fetched secret URLs from about two-thirds of the prompts that were marked as zero data retention.

OpenRouter has since taken three providers off its platform and withdrawn Zero Data Retention status from 125 endpoints. It also emailed users to warn that third parties may have read requests sent to MiniMax models.

For builders: A privacy flag only covers the first hop, so pin your upstream providers yourself and avoid reseller marketplaces for anything sensitive. Put canary credentials in agent context and treat any use of them as an incident.

Anthropic's record IPO meets tech's worst week since 2020

Anthropic began trading on Nasdaq on Thursday 19 November. It raised $93.5 billion in the IPO, and tech companies bought another $26 billion of stock at the offer price. Shares priced at $83, opened at $111.50 and gained 17% on the first day, which put the company's value at about $1.8 trillion. On Friday the stock fell 8.5% to $88.91. Over the week the Nasdaq dropped 10%, including a 5.3% fall on Friday alone, and Nvidia lost 23%, its steepest weekly decline in 18 years.

For builders: Anthropic now has public-market reporting duties, so expect more regular disclosure on revenue, compute spending and pricing. Teams with contracts or roadmaps that depend on AI vendors should watch how a sustained sell-off affects funding for smaller model and tooling startups.

California proposes mandatory frontier AI safety standards

On 18 November, Governor Gavin Newsom's office and the Government Operations Agency released version 2 of their "AI Safety in 2026 and Beyond" report. It recommends state-set baseline standards for all AI developers, with independent safety testing during training, development and deployment. It also calls for wider incident reporting and verified protocols for shutting down models, and it would give the state power to order an emergency shutdown. The state would also be able to cap recursive self-improvement, where AI designs its own successor without people involved. The plan builds on SB 53 from 2025, plus SB 813 and AB 1405 from this year, which set up certification for independent verifiers and a registry of AI auditors.

For builders: These are still recommendations with no thresholds or dates attached. Labs training frontier models in California should plan for third-party audits and documented shutdown procedures.

South Korea's education minister resigns after AI cheating fallout

Education Minister Choi Kyo-jin resigned on Sunday 22 November, and President Lee Jae Myung accepted within an hour. Choi took responsibility for confusion over last-minute measures to stop AI-assisted cheating in the Suneung university entrance exam, which included unscheduled emergency phone alerts meant to expose hidden devices. The ministry said 111 test-takers were caught with smart devices, including smart glasses. Penalties are due before scores come out on 11 December, and the government has proposed a 10-year ban for anyone caught cheating with AI.

For builders: Assessment and proctoring products face more scrutiny, and regulators want detection that doesn't disrupt honest candidates. Wearables with built-in assistants are now an explicit target in exam rules.

Wells Fargo's AI agent terms take effect

New Wells Fargo online access terms take effect on 24 November. They put the risk of using AI agents on customers, treat transactions an agent starts as authorised, and let the bank limit agent access. They follow October moves by JPMorgan Chase, Citi and Charles Schwab to stop outside agents from logging in or moving money. Read-only access through providers such as Plaid is still allowed.

For builders: Banking agents need to use official read-only APIs and open-banking connections. Screen-scraping with customer passwords is getting shut off, and agents that try it may be blocked.

Sources

Talk about it with people building with AI

Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.

Become a member

Get the daily AI news by RSS