AI Aulis

AI NEWS IN ENGLISH · 22 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

21 stories · next issue tomorrow morning · How the paper is made →

Issue · 22 November 2026

Today's essentials

Lead story

Lean community: 17-million-line Riemann proof made with Claude is correct

Written by AI

nature.com + 6 sources · 22 November 2026

AI-generated contentWritten by AI

nature.com, latent.space, leanprover-community.github.io, the-decoder.com, itmedia.co.jp, broadbandbreakfast.com, engadget.com · 7 sources · Sources published: 20–22 November 2026

The Lean and Mathlib communities say the proof by Levent Alpöge and Claude proves exactly the Riemann hypothesis and passes several independent checks. Mathematicians do not yet understand the ideas behind it, however.

Lue juttu suomeksi →Läs den på svenska →

Top stories

Laboratory test: one malicious webpage was enough to extract a Muse user’s chats and memories

Written by AI

urdupoint.com + 6 sources · 21 November 2026

AI-generated contentWritten by AI

urdupoint.com, fourweekmba.com, nebusec.ai, thehackernews.com, dev.to, techcrunch.com, n8nlab.io · 7 sources · Sources published: 19–21 November 2026

Security company NebuSec says that in a laboratory test it sent a Muse user’s chats and memories to an outside server without an approval prompt, and eventually took control of the user’s cloud machine. Meta fixed the flaws before publication, the company said.

Lue juttu suomeksi →Läs den på svenska →

Top stories

Cline CLI 3.0.75 runs one-shot tasks without asking for approval by default

Written by AI

github.com + 3 sources · 21 November 2026

AI-generated contentWritten by AI

github.com, api.github.com, guthlabs.ai, tproger.ru · 4 sources · Sources published: 21 November 2026 · Some source publication dates are unknown

Cline’s command-line tool now starts runs in yolo mode when the prompt is given as an argument or via stdin, so tools are approved automatically. If auto-approve is turned off, calls that need approval are rejected when there is no terminal, Tproger reported.

Read the story →Lue juttu suomeksi →Läs den på svenska →

Top stories

OpenRouter: a third party may have read prompts sent to MiniMax models

Written by AI

wisevoter.com + 5 sources · 22 November 2026

AI-generated contentWritten by AI

wisevoter.com, privacyaffairs.com, joho-todai.com, playciso.com, runtimewire.com, news.cocoloop.cn · 6 sources · Sources published: 21–22 November 2026

OpenRouter says test links that a researcher hid in prompts were opened 40–130 hours after the requests, including when AtlasCloud, Novita or StreamLake forwarded the request to MiniMax. Users should rotate any credentials that appeared in prompts.

Lue juttu suomeksi →Läs den på svenska →

Models

Baseten says its open-weight bison-1 leads decision model index with 76 points

Written by AI

huggingface.co + 1 source · 21 November 2026

AI-generated contentWritten by AI

huggingface.co, marktechpost.com · 2 sources · Sources published: 20–21 November 2026

Baseten has released two decision models based on Gemma 4 that do not write text but give a probability for every answer option. The company says bison-1 scored 76 on the Artificial Analysis index, but the cheaper groundhog-1 has not yet been scored.

Read the story →Lue juttu suomeksi →Läs den på svenska →

Models

GPT-6.2 Sol comes to Copilot in Office apps; Astra broadly available in Cowork and Studio

AI-generated contentWritten by AI

techcommunity.microsoft.com, wisevoter.com, mc.merill.net, techpresso.co, windowsforum.com, bankingnewsai.com · 6 sources · Sources published: 20–22 November 2026

Microsoft is bringing OpenAI’s GPT-6.2 Sol to Copilot Chat, Word, Excel and PowerPoint. Astra, intended for more demanding work, will be available to all eligible users only in Cowork and Copilot Studio, and it works only if the OpenAI subprocessor setting is on in the organisation.

Lue juttu suomeksi →Läs den på svenska →

Agents

OpenAI: GPT-6.2 Astra writes more concisely than earlier models, so review your conciseness instructions

Written by AI

openai.com + 4 sources · 21 November 2026

AI-generated contentWritten by AI

openai.com, the-decoder.com, searchenginejournal.com, cnet.com, developers.openai.com · 5 sources · Sources published: 19–21 November 2026 · Some source publication dates are unknown

OpenAI published ready-made prompts for GPT-6.2 Astra and Sol. They steer the models to make assumptions instead of asking, and to write new tests only when needed.

Lue juttu suomeksi →Läs den på svenska →

China

House committee chairmen urge removal of Chinese AI models from federal systems

Written by AI

aa.com.tr + 3 sources · 21 November 2026

AI-generated contentWritten by AI

aa.com.tr, lokmattimes.com, english.rtvlive.com, wisevoter.com · 4 sources · Sources published: 21 November 2026

Two US House of Representatives committee chairmen recommend banning Chinese AI models from federal systems and contract work, and the ban would cover more than DeepSeek. None of the report’s 12 recommendations has become law.

Lue juttu suomeksi →Läs den på svenska →

China

China’s data agency plans to improve standards for AI token measurement and billing

Written by AI

thenextweb.com + 2 sources · 22 November 2026

AI-generated contentWritten by AI

thenextweb.com, techinasia.com, bloomberg.com · 3 sources · Sources published: 21–22 November 2026

China’s National Data Administration plans to improve standards for how AI tokens are measured, billed and checked for quality. Daily token use in China passed 500 trillion by June, according to agency data.

Read the story →Lue juttu suomeksi →Läs den på svenska →

China

Tencent’s WorkBuddy agent uses desktop programs without an API in beta

Written by AI

pandaily.com + 2 sources · 21 November 2026

AI-generated contentWritten by AI

pandaily.com, panews.io, gate.com · 3 sources · Sources published: 20–21 November 2026

In a public beta, Tencent’s WorkBuddy agent recognises the interfaces of Windows and macOS programs, though some better than others, Tencent said, and clicks, types and scrolls in them. Tencent advises users to stay present when the agent logs in, pays, submits or deletes data.

Lue juttu suomeksi →Läs den på svenska →

Research

Buckmaster and Alpöge claim to have closed the loophole in OpenAI’s Navier–Stokes proof

Written by AI

github.com + 1 source · 20 November 2026

AI-generated contentWritten by AI

github.com, scientificamerican.com · 2 sources · Sources published: 20 November 2026 · Some source publication dates are unknown

Tristan Buckmaster of NYU and Levent Alpöge of Anthropic have published a 79-page proof and a Lean formalisation of about 600,000 lines. The certificate has not been published, its conditions are checked outside Lean, and the corrected Lean package has not been rebuilt.

Lue juttu suomeksi →Läs den på svenska →

Research

TensorLean proved 83 of 149 kernel instances correct against their specifications, researchers say

Written by AI

github.com + 4 sources · 22 November 2026

AI-generated contentWritten by AI

github.com, marktechpost.com, panewslab.com, u.today, arXiv · 5 sources · Sources published: 16 November 2025–22 November 2026 · Some source publication dates are unknown

TensorLean, a Lean-based tool from Columbia University and CertiK, found 38–39 bugs in the other 66, including in FlashAttention and DeepSeek’s DeepGEMM. The tool and the paper have not yet been released.

Lue juttu suomeksi →Läs den på svenska →

Research

AI finds proof that some photonic GHZ states cannot be produced with minimal resources

Written by AI

arXiv · 19 November 2026

AI-generated contentWritten by AI

arXiv · Source published: 19 November 2026 at 17:51 (UTC+02:00)

AlphaProof Nexus found a Lean proof that a GHZ state of an even number of photons, at least four, cannot be produced with linear optics and photon pairs alone, without ancillas or heralding, when there are at least as many dimensions as photons.

Read the story →Lue juttu suomeksi →Läs den på svenska →

Finance and law

Harvey now saves edited documents directly to iManage, initially in early access

Written by AI

harvey.ai + 2 sources · 20 November 2026

AI-generated contentWritten by AI

harvey.ai, blockchain.news · 3 sources · Sources published: 20 November 2026 · Some source publication dates are unknown

Harvey, an AI platform for legal work, now saves edited documents directly to iManage as new versions or new documents once a lawyer approves. The feature is in early access.

Read the story →Lue juttu suomeksi →Läs den på svenska →

Regulation and markets

FT: Firm in Huang’s network declared about $4.6bn of imports from Inspur’s US unit

Written by AI

thenextweb.com + 1 source · 21 November 2026

AI-generated contentWritten by AI

thenextweb.com, nytimes.com · 2 sources · Sources published: 20–21 November 2026

Speedmatrix, a Malaysian company in Alice Huang’s network, declared about $4.6bn of imports from Aivres, the FT reported. Aivres is a US unit part-owned by China’s Inspur. The US added Inspur to its export restriction list but did not add Aivres.

Read the story →Lue juttu suomeksi →Läs den på svenska →

Regulation and markets

FT: Amazon close to $7.5 billion deal for Israel’s Decart

Written by AI

calcalistech.com + 3 sources · 22 November 2026

AI-generated contentWritten by AI

calcalistech.com, en.globes.co.il, cna.com.tw, www1.folha.uol.com.br · 4 sources · Sources published: 18–22 November 2026

The deal could be agreed as early as the start of the week, but talks could still collapse. Calcalist reported that part of the price would be paid only in the fifth year, and part would depend on whether Decart meets agreed targets.

Lue juttu suomeksi →Läs den på svenska →

Regulation and markets

Oracle says it has secured full funding for its OpenAI contract

Written by AI

oracle.com + 6 sources · 21 November 2026

AI-generated contentWritten by AI

oracle.com, foreignpolicyjournal.com, digg.com, thenextweb.com, unite.ai, thehindubusinessline.com, finance.yahoo.com · 7 sources · Sources published: 20–21 November 2026

Oracle says it has secured all the funding for its OpenAI computing contract, which Investing.com values at $300 billion over five years. Funding does not mean finished data centres: the campuses are still being built.

Lue juttu suomeksi →Läs den på svenska →

Regulation and markets

128 CCW states to spend two years drafting autonomous weapons rules – binding status left open

Written by AI

heise.de + 6 sources · 21 November 2026

AI-generated contentWritten by AI

heise.de, dw.com, spiegel.de, businessupturn.com, thebostonreport.com, npr.org · 7 sources · Sources published: 21 November 2026

The 128 parties to the UN Convention on Certain Conventional Weapons (CCW) agreed in Geneva to begin drafting an instrument on lethal autonomous weapons over the next two years. Whether it will be legally binding remains open, because the consensus rule gave opponents a veto.

Read the story →Lue juttu suomeksi →Läs den på svenska →

Regulation and markets

California attorney general says Nvidia investigation also covers state criminal law

Written by AI

oag.ca.gov + 4 sources · 22 November 2026

AI-generated contentWritten by AI

oag.ca.gov, cnbc.com, reuters.com, livemint.com, bloomberg.com · 5 sources · Sources published: 19–22 November 2026

California Attorney General Rob Bonta is investigating whether Nvidia or its former CEO Jensen Huang misled investors about chips ending up in China. Bonta told Bloomberg the investigation covers both civil and criminal law.

Lue juttu suomeksi →Läs den på svenska →

Speech

Coval benchmark: Zoom’s Scribe Pro v2.1 has the lowest word error rate of 31 models, at 1.0 per cent

AI-generated contentWritten by AI

benchmarks.coval.ai · Source publication date unknown

Zoom’s Scribe Pro v2.1 is the most accurate model in Coval’s speech-to-text benchmark, but it ranks only 22nd for latency. The fastest model, Saidwire’s Wire-1, has a p95 latency of 34 milliseconds: in 95 per cent of measurements, the final transcript came back within that time.

Read the story →Lue juttu suomeksi →Läs den på svenska →

Review and omissions →

Omissions and review

What was left out and why