A Claude-assisted Riemann Hypothesis proof passes formal checks
On Friday 20 November, Anthropic mathematician Levent Alpöge posted a note saying Claude had helped him prove the Riemann Hypothesis over about six weeks. He released the argument as a Lean library of roughly 17 million lines. On 21 November the Lean FRO and the Mathlib Initiative said they had "full confidence" the proof is correct. Dozens of people rebuilt the library. Two separate tools confirmed that the proved theorem matches Mathlib's definition, and several independent checkers, including Lean's own kernel, accepted the proof.
Bernhard Riemann posed the problem in 1859. It's one of the Clay Mathematics Institute's seven Millennium Prize problems, each worth US$1 million. The verifying groups say mathematicians still have to explain the new ideas inside the proof.
What it means for NZ organisations: Machine-checked verification is becoming the way to trust AI on hard technical work. Where you use AI for code or analysis, add automated tests and checks so you don't have to rely on the model's own confidence.
Attackers sat inside AI inference providers for months
On 19 November Dragos published the results of an experiment. Between July and November it sent about two million prompts containing trackable fake credentials. It found attackers inside at least three AI hosting providers, the earliest from mid-July. They took a couple of hundred bait AWS keys and often tested them within hours. They made tens of thousands of API calls and tried to add IAM users and Bedrock keys. Dragos also found that some marketplace resellers kept prompts through ordinary logging.
On 22 November OpenRouter said it had removed three providers and taken 125 endpoints off its Zero Data Retention list. It said its own systems weren't breached. Separately, Together AI disclosed on 21 November that an AI coding agent had committed a credential to a public GitHub repository. Someone used it to download samples of customer prompts from June and July.
What it means for NZ organisations: Treat anything sent through low-cost hosts or routers as readable by others. Keep credentials out of prompts, contract with providers directly, and consider planting fake credentials so you can spot misuse.
California proposes mandatory frontier AI safety rules
On 18 November Governor Gavin Newsom released recommendations from the state's Government Operations Agency. They include:
- minimum safety standards set by the state
- independent testing across training, development and deployment
- wider incident reporting
- verified shutdown procedures, with state power to order an emergency shutdown
- authority to cap recursive self-improvement
Newsom asked Congress and President Trump to adopt the framework or use it as a floor. He also told agencies to fast-track two laws: SB 813, which certifies independent verification organisations, and AB 1405, which creates a state registry of AI auditors.
What it means for NZ organisations: Most frontier models used here come from California developers, so more independent audit material is likely. Ask vendors for it during procurement.
US banks shut out third-party AI agents
Wired reported on 22 November that big US banks are restricting outside AI agents. Wells Fargo's updated online access terms take effect on Tuesday 24 November. They say any use of an AI agent is at the customer's own risk. In late October JPMorgan Chase banned third-party agents that log in with customer credentials. Bank of America warned customers not to share logins with agents, and Citibank and PNC added similar warnings. Amazon has blocked Meta's Muse agent from making purchases, and United is stopping bots from booking flights.
What it means for NZ organisations: If you're building agents that act on customer accounts, check the platform's terms now and plan to use official APIs where they exist.
Signal blocks AI assistants from reading chats
In a 20 November post, Signal president Meredith Whittaker said the app now blocks screenshots on iOS 27. On Android, it turns on its Screen security setting when it detects an AI assistant reading the screen. Signal can't reliably do this on macOS. There, it suggests adding Signal to the Excluded Apps list in the Siri settings. Users can turn the protection off. Signal asked Apple, Google and Microsoft to give developers proper opt-out tools.
What it means for NZ organisations: Assistants built into the operating system can see whatever's on screen. Review which assistant features are enabled on managed devices that handle confidential work.
South Korea's education minister resigns over AI cheating
Choi Kyo-jin resigned on Sunday 22 November. The trigger was a confession from the top scorer in a national tax accountancy exam. He said he had photographed the questions with a hidden camera, sent them to an AI system and read the answers from a phone.
Before the university entrance exam on 19 November, officials took several steps:
- doubled the number of supervisors
- blocked public wifi near exam centres
- sent emergency alerts to every phone to find hidden devices
They identified 111 cases of misconduct involving devices, nearly half of them through the alerts. President Lee Jae Myung accepted Choi's resignation.
What it means for NZ organisations: Remote or lightly supervised assessments are exposed. Training providers and employers running certifications should review how they check identity and devices.
UBS raises its AI investment forecast
A UBS report covered on 22 November expects global AI investment to rise from US$1 trillion in 2026 to US$1.4 trillion in 2027 and US$1.5 trillion in 2028. It says about 80% of 2027 spending will come from Alphabet, Amazon, Meta, Microsoft and SpaceX. UBS warned that returns are uncertain. It named power, transmission and data centre shortages among the risks.
What it means for NZ organisations: Spending by the largest cloud providers keeps capacity growing, but prices may move. Build flexibility into multi-year cloud and AI contracts.
Sources
- Lean verification of the Riemann Hypothesis (Lean community blog)
- Three breaches, endless leaks: Your AI supply chain needs canaries (Dragos)
- Data processing incidents on inference providers (OpenRouter)
- Security event: prompt leakage for requests in June and July of 2026 (Together AI)
- Governor Newsom's AI safety framework next steps (Office of the Governor of California)
- Protecting your privacy (Signal blog)
- Two groups verify mathematician's claim that Claude helped prove mathematics problem (The Guardian)
- South Korean minister quits after AI cheating scandal engulfs university exams (The Guardian)
- Banks Have Had It Up to Here With Your AI Agent (Wired)
- Signal restricts mobile screenshots as it confronts existential threat of AI surveillance (TechCrunch)
- OpenRouter Cuts Three Providers After AI Prompt Exposure Findings (Unite AI)
- Together AI discloses prompt leak after coding agent exposed credential (Unite AI)
- AI investment to rise to USD 1.4 trillion in 2027, but returns remain key test: UBS (Mint)
- AI News Digest: 22 November 2026 (AIE News)
Want to know which of these changes matter for your organisation? Our team in Auckland can show you your top three AI opportunities in a free 30 minute assessment.
Get my free assessment How we help