Today's Headlines
- An Anthropic mathematician says he proved the Riemann Hypothesis with Claude
- OpenAI details its defenses against adversarial distillation
- Signal blocks AI agents built into iOS and Android from reading its screen
- Anthropic closes its IPO, including the underwriters' full option
- Thinking Machines Lab releases the open-weights Inkling 2 Mini
A claim about one of mathematics' hardest problems leads the day. OpenAI and Signal each moved to protect what sits behind their products: a model's capabilities and an app's messages.
Today's Top Three
An Anthropic mathematician says he proved the Riemann Hypothesis with Claude
Levent Alpöge, a mathematician at Anthropic, said on November 20 in a document on his website that he has found a proof of the Riemann Hypothesis working with Claude.
The Riemann Hypothesis, proposed in 1859, concerns the distribution of prime numbers. It is one of the seven Millennium Prize Problems, for each of which the Clay Mathematics Institute offers $1 million.
Alpöge writes that the same method also resolves the Generalized Riemann Hypothesis. He says he and Claude have formalized the argument in the Lean proof assistant, and he has released only the Lean files for now. He says he intends to write the argument up properly.
Independent verification is still to come. Under the Clay Mathematics Institute's rules, a proposed solution is considered only after publication in a qualifying outlet, at least two years after that publication, and general acceptance in the mathematics community.
TechCrunch reports that a separate proposed proof of the Generalized Riemann Hypothesis by another researcher was posted the same day, described by those who released it as entirely human work, with no AI.
- Document by Levent Alpöge (alpo.ge, PDF)
- Anthropic mathematician claims to have solved prestigious Millennium Prize problem with Claude (The Verge)
- AI claimed a solution to the 167-year-old Riemann Hypothesis — and a human did, too (TechCrunch)
- Rules for the Millennium Prize Problems (Clay Mathematics Institute)
OpenAI details its defenses against adversarial distillation
OpenAI on November 20 described how it detects and disrupts adversarial distillation, in which someone uses a model's outputs without authorization to train another model with similar capabilities.
Its main focus is protected reasoning, the chain of thought a model generates on the way to an answer. OpenAI says these traces carry information the final answer leaves out, which makes them especially valuable training data. It is most concerned about dual-use fields such as cybersecurity and biology, where a copied capability could end up in a model without comparable safeguards.
In OpenAI's internal adaptive red-teaming, the average success rate of attacks extracting protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra with its original safeguards to 0.2% on GPT-6.2 Astra with updated safeguards. For reasoning from the current request, the rate fell from 22.9% to 0.9%, the company says.
The defenses work at two levels. Model training, classifiers and API controls protect individual requests, while investigations look for coordinated extraction campaigns across accounts. OpenAI says its weekly enforcement actions against such campaigns have ranged from thousands to tens of thousands of accounts, rising significantly this past summer.
OpenAI says it supports legislation that would give the US government more tools to identify and sanction foreign actors that extract capabilities from American models. It has told Congress it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act.
Signal blocks AI agents built into iOS and Android from reading its screen
Signal said on November 20 that it is restricting AI agents built into mobile operating systems from capturing what appears on its screen.
On iOS 27 the app now blocks screenshots, and on Android it blocks them when it can detect an AI assistant trying to capture the screen. Signal says agentic features such as Google's Gemini and Apple's Siri, built deep into the operating systems, are undermining its ability to guarantee privacy on its own.
Users can turn the protection off in settings, which Signal says may expose their own and their contacts' messages to these agents.
Signal took a similar step against Microsoft's Recall on Windows 11 last year. This time it called on Apple, Google and Microsoft to give developers a simple, documented way to opt their apps out of AI data access, from screen reading to file access, and to publish clear documentation of how their AI agents work.
- Signal is Serious About Protecting Your Privacy (Signal)
- Signal is blocking Gemini and Siri AI from reading your screen (The Verge)
Other developments
Policy and Safety
- President Trump said at a press conference on November 20 that he would put a "tremendous tax" on companies that build data centers outside the US, The Verge reports. The outlet says he appeared to be floating the idea publicly for the first time.
Trump threatens ‘tremendous tax’ for moving AI data centers outside the US (The Verge)
Models and APIs
- Thinking Machines Lab released Inkling 2 Mini, an open-weights model that runs its middle layers three times for every token. It is a mixture-of-experts model with 26.4 billion total and 3.4 billion active parameters and a memory footprint of about 20GB, small enough for a laptop. The company says it leads open-weights models under 100 billion total parameters on benchmarks such as DeepSWE v1.1 (48.2%).
Introducing Inkling 2 Mini (Thinking Machines Lab)
Business
- Anthropic closed its initial public offering on November 20. The underwriters exercised in full their option to buy 147 million additional shares, bringing the offering to 1,127,000,000 shares at $83 for gross proceeds of about $93.5 billion, about $91.6 billion of which went to Anthropic. It also sold about $26.0 billion of stock at the same price in a concurrent private placement to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners, for a combined total of about $119.5 billion.
Anthropic announces closing of initial public offering (Anthropic)
Source: AI news inbox (collected November 21, 2026; 6 items, 2 primary and 4 secondary), selected by the editors.