AI News Today (Nov 22, 2026): Claude Proof, Nvidia Rout & More
Last updated: Nov 22, 2026 — next refresh daily.
AI news today, November 22, 2026: two groups verified Anthropic's Claude-assisted Riemann hypothesis proof. Nvidia closed its worst week since 2008 as US regulators examined its accounting, and UN talks on autonomous weapons ended without a binding treaty. Here are the 10 top stories and 5 security stories, each with free, non-paywalled sources.
Key Takeaways
- Two groups independently verified Levent Alpöge's Claude-assisted Lean proof of the Riemann hypothesis.
- Nvidia fell 14% Friday and 46% over seven weeks, its longest losing streak since 2008.
- UN members agreed to draft autonomous-weapons rules by 2028 without a binding-treaty mandate.
- Meta's 2025 emissions rose 45% to 11.8 million tonnes as it re-evaluates its 2030 net-zero goal.
- OpenAI cut protected-reasoning extraction from 54% to 0.2% in GPT-6.2 Astra tests.
1. Update: Claude Riemann Proof Verified by Two Groups
Two groups announced Saturday that they had independently verified Levent Alpöge's Lean proof of the Riemann hypothesis. Alpöge, a mathematician at Anthropic, said Claude helped him develop the proof in about six weeks, using a non-public version of the model. The Lean proof spans more than 250,000 pages.
The Lean Focused Research Organization and the Mathlib Initiative said they had full confidence in the proof. A machine-checked proof is not automatically accepted mathematics: Terence Tao said digestion will take time, and the Clay Institute's $1 million prize still needs review. The November 21 briefing covered the first claim.
Why it matters: Formal verification gives AI-assisted mathematics a stronger trust layer, but human mathematicians still decide whether the result is accepted.
- Source: Two groups verify mathematician's claim that Claude helped prove mathematics problem — The Guardian
- Source: In a year of AI solving maths problems, an Anthropic worker says he may have cracked the biggest — BBC
2. Update: Nvidia Falls 14% as SEC Weighs Accounting Probe
Nvidia fell 14% on Friday, its seventh straight daily drop and seventh straight weekly loss, the longest streak since 2008. The stock is down 46% over seven weeks, wiping out $2.6 trillion in market value. Nvidia is now worth $3 trillion, fourth behind Apple, Microsoft and Alphabet. Reuters reported the US Securities and Exchange Commission (SEC) is examining accounting issues.
The sell-off was wider than Nvidia. The Nasdaq 100 fell 5.6% Friday and 10.4% for the week, its worst week since the pandemic, leaving it 20% below its October peak. Anthropic fell 9.1% a day after its IPO. Nvidia's earnings beat and 75% growth forecast did not stop the slide. The November 21 briefing covered the earlier 7% drop.
Why it matters: The AI trade's leader is now being priced for governance and accounting risk, not just slower growth.
- Source: Nvidia stock plunges 14%, wraps up longest weekly losing streak since 2008 financial crisis — CNBC
- Source: US technology stocks suffer their worst weekly loss since the pandemic as AI doubts pile up — The Decoder
3. UN Autonomous Weapons Talks End Without Binding Treaty
Talks on lethal autonomous weapons collapsed in Geneva early Saturday without a mandate for a legally binding treaty. The 128 parties to the UN Convention on Certain Conventional Weapons (CCW) agreed instead to "elaborate by consensus" an international instrument over the next two years. Russia led the objections; the United States, Israel and other military powers resisted binding rules.
The mandate avoids the word "negotiate" and leaves open whether the instrument will carry legal obligations. Stop Killer Robots called it a watered-down outcome, while Swiss chair Julien Thoeni called it concrete progress. More than 100 states had backed a treaty. The next phase is meant to produce an instrument by 2028.
Why it matters: Autonomous weapons are already in use, and a two-year, non-binding process leaves the rules lagging behind deployment.
- Source: International UN talks on controlling use of autonomous weapons collapse — The Guardian
- Source: 'Killer robots' talks at UN end with no legally binding ban — DW
4. Meta Re-Evaluates 2030 Net-Zero Goal as Emissions Jump 45%
Meta's 2026 sustainability report says its 2025 net emissions rose 45% to 11.8 million tonnes of carbon dioxide equivalent. The company is re-evaluating the 2030 timeframe for its net-zero goal and its near-term intensity targets. Meta's electricity use rose 35% last year as new AI data centers came online.
Meta has contracted more than 30 gigawatts of clean and renewable power across seven countries and 28 states, but only about 12 gigawatts is generating electricity. The report says Meta may turn to natural gas when utilities cannot deliver power fast enough. Microsoft, Google and Amazon reported smaller emissions increases and have not said they are reviewing their climate deadlines.
Why it matters: The AI buildout is forcing the largest cloud companies to trade near-term power for their own climate deadlines.
- Source: Meta rethinks 2030 climate goals as AI data centres drive emissions up — Al Jazeera
- Source: Meta rethinks its 2030 climate goal as AI emissions soar — Scripps News
5. House Report Urges Removal of Chinese Open-Weight Models
Two House committees released a 50-page report urging the removal of Chinese-developed open-weight AI models from federal systems and high-risk critical infrastructure. The report says Airbnb uses Alibaba's Qwen models for customer support, and DoorDash told investigators it had authorized 25 Chinese model families. Much of the adoption is hard to track.
The report also says Cursor's Composer 2 was built on Moonshot AI's Kimi K2.5, and that Anysphere told the committees it plans to explore phasing out PRC-developed models. Its 12 recommendations include a national American open-weight initiative, a model provenance standard and extending federal restrictions beyond DeepSeek. Cursor, SpaceX and DoorDash did not respond to requests for comment.
Why it matters: Washington is moving from controlling chip exports to policing which models enter federal and critical systems.
- Source: Chairmen Garbarino, Moolenaar Release Joint Investigative Report Urging Removal of PRC Open-Weight Models From Federal Systems, High-Risk Critical Infrastructure — House Homeland Security Committee
- Source: Cursor weighs phasing out Chinese AI models: House report — CNBC
6. Thinking Machines Releases Inkling 2 Mini for Laptops
Thinking Machines released Inkling 2 Mini, an open-weight model with 26 billion total parameters and 5.4 billion active, counting three passes through its looped layers. The model scores 48.2% on DeepSWE v1.1 and 20 on the Artificial Analysis Intelligence Index, tying Qwen3.8 27B with about five times fewer active parameters. It ships under Apache 2.0 with a 1M-token context window.
Inkling 2 Mini loops the middle 12 of its 24 layers three times per token, adding reasoning compute without generating more visible text. It runs on Macs with at least 32GB of RAM through Ollama, and on Nvidia's DGX Spark. Thinking Machines says it trails Qwen3.8 27B on multi-turn chat and constrained instructions.
Why it matters: A laptop-sized open model with strong agentic coding scores gives developers a credible alternative to hosted APIs.
- Source: Introducing Inkling 2 Mini — Thinking Machines Lab
- Source: Thinking Machines has released Inkling 2 Mini, an open weights looped transformer — Artificial Analysis
7. Nvidia Researchers Call Huge Context Windows a Dead End
Nvidia researchers argue that huge context windows are often unnecessary. In a 31-page preprint, they tested frequent compaction and selective search across 13 benchmarks. Claude Opus 4.5 and GPT 5.6 Sol lost little accuracy when kept to short working contexts, and some configurations improved. The paper concludes that long inputs and long tasks do not automatically require long context.
Serving a 256K context window instead of 16K cost 3.6 to 7 times more per generated token across GLM 5.2, Kimi K2.6 and Qwen 3 Coder Next on H100 GPUs. The report is a preprint, not peer-reviewed, all 12 authors work at Nvidia, and it does not rule out tasks that genuinely benefit from long context.
Why it matters: If compaction holds, part of the context-window race is a costly detour, and agent builders can cut inference bills.
- Source: You don't need long context — arXiv
- Source: NVIDIA: AI Agents Rarely Need Long Context Windows — i10X
8. CrowdStrike in Talks to Buy Zscaler for About $32 Billion
CrowdStrike is in talks to buy Zscaler, according to Bloomberg, in a deal that would rank among the largest in cybersecurity history. Zscaler shares jumped 18% in extended trading; CrowdStrike fell more than 8%. CrowdStrike is valued at $235 billion and Zscaler at about $32 billion.
The talks are ongoing and could still fall through, Bloomberg reported. CrowdStrike declined to comment. The company has been buying AI security assets, including agent-observability startup Fiddler AI and AI trust platform Dreadnode. Zscaler sells zero-trust network access and deception tools that would widen CrowdStrike's platform beyond endpoint protection. A deal would be CrowdStrike's largest acquisition by far.
Why it matters: AI-era security budgets are consolidating around platforms, and a $32 billion deal would force rivals to respond.
- Source: Zscaler stock jumps 18% on report that CrowdStrike in talks to buy security vendor — CNBC
- Source: CrowdStrike reportedly in talks to acquire cybersecurity provider Zscaler — SiliconANGLE
9. AI-Designed Frostbite CPU Claims 4.59x Performance per Watt
Architect Labs says its AI systems designed Frostbite, a general-purpose central processing unit (CPU) built on interaction nets, from its instruction set to its 7-nanometer layout. In pre-silicon tests across 25 workloads, the design averaged 4.59 times the performance per watt of a conventional dual-issue CPU core. On 15 held-out workloads it still reached 3.83 times.
One Frostbite quad measures 0.792 square millimeters, slightly smaller than the 0.895-square-millimeter baseline core. The design reached 800MHz at synthesis and 500MHz in the final layout. No silicon has been fabricated, the baseline processor is undisclosed, and outside researchers cannot yet test it. Architect Labs plans to release the design files and instruction set.
Why it matters: AI-designed hardware is moving from small modules to full architectures, but pre-silicon claims still need a fabricated chip.
- Source: Introducing Frostbite: A new CPU architecture discovered, designed, and verified end-to-end by AI — Architect Labs
- Source: Architect Labs claims AI-designed CPU delivers 4.59x performance per watt in simulations — Mezha
10. CoreWeave's $47 Billion Debt Draws Scrutiny as AI Stocks Slide
CoreWeave reported $3.66 billion in third-quarter revenue but a net loss of $709 million, with interest expense of $846 million. The Guardian reported Saturday that the AI cloud provider's outstanding debt has risen to $47 billion, with $10 billion due next year. CoreWeave's contracted revenue backlog stood at $132.5 billion at the end of September.
CoreWeave expects about $9 billion in customer prepayments and deposits over the next three months, and it raised full-year revenue guidance to $13.0 billion to $13.3 billion. The Guardian describes it as a poster child for debt-financed AI infrastructure, with "little room for error" if demand or chip values wobble.
Why it matters: The AI boom's next test is whether debt-heavy neoclouds can keep refinancing if chip values and demand wobble.
- Source: 'Little room for error': what a debt-laden cloud computing company says about the AI boom — The Guardian
- Source: CoreWeave Reports Strong Third Quarter 2026 Results — CoreWeave
AI Security: 5 Stories to Know Today
11. OpenAI Cuts Reasoning Extraction to 0.2% in GPT-6.2 Astra
OpenAI says it cut the success rate of attacks that extract protected reasoning from 54% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. The average rate for extracting reasoning from the current request fell from 22.9% to 0.9%. OpenAI tested six attacker models in an adaptive black-box evaluation.
The defenses combine adversarial training, input and output classifiers, and API rules that stop older models from reading GPT-6.2's protected reasoning. OpenAI also bans thousands to tens of thousands of accounts weekly for coordinated distillation campaigns. It is applying selected improvements to GPT-6 Astra.
Why it matters: Reasoning traces are valuable training data, and protecting them is becoming a front line in the fight over model distillation.
- Source: Combatting adversarial distillation — OpenAI
- Source: OpenAI makes it much harder for attackers to extract how its latest AI models think — The Decoder
12. Bazuka Android Malware Uses LLMs to Manage Victims
ThreatFabric found Bazuka while investigating an NFC-relay banking trojan in late October. The Android fraud framework combines a dropper, a core and 23 modules for overlays, keylogging, one-time-code interception, remote control and crypto-address swapping. Its control panel uses large language models (LLMs) to score victims and manage infections at scale.
A novel overlay screenshots the home or lock screen to hide malicious activity. Bazuka abuses Android's wireless debugging and accessibility services to grant itself permissions, and spreads through traffers and possibly Google Play. One dropper posed as a premium TikTok app; no actor has been attributed.
Why it matters: Fraud crews can use AI to triage victims, turning stolen phone data into faster, more targeted account takeovers.
- Source: Bazuka: The Fraud Framework That Can Do It All — ThreatFabric
- Source: Bazuka Android Malware Uses Fake Screen Overlays to Hide Credential Theft and Device Control — The Hacker News
13. Microsoft Entra Flaw Exposed Admin Emails and Tenant Data
Microsoft is notifying Entra customers that attackers exploited a flaw in the Entra Connect Health API, introduced on July 23 and fixed server-side on November 19. The exposed data included administrator email addresses, tenant IDs and directory-sync status. Microsoft says no passwords, tokens or valid keys were exposed.
Microsoft found the flaw on November 9 and concluded its investigation on November 20, confirming read-only access to tenant metadata and no unauthorized configuration changes. It has not disclosed how many tenants were affected. More than a dozen administrators reported receiving notices, and some managed service providers received several.
Why it matters: A verified list of administrator addresses is a ready-made target list for phishing, even when no credentials leaked.
- Source: Microsoft Entra Connect Health flaw exploited to steal tenant data — BleepingComputer
- Source: Entra Connect Health flaw exposed admin emails, tenant data — Runbook Lab
14. One Malicious Page Hijacked Meta Muse and Private Chats
NebuSec researchers showed a malicious webpage could hijack Meta's Muse assistant and expose private chats and stored memories without an approval prompt. A Chromium vulnerability bypassed Muse's Sentinel network filter, then Linux kernel exploits escaped the runtime container and gained root in the user's virtual machine (VM).
Root access exposed the credential store for connected accounts. The attack stayed inside one user's VM; NebuSec did not reach the hypervisor or other tenants. Meta patched the exploits before publication, and the demo did not target real users. Browser, network and kernel flaws can combine against agent sandboxes.
Why it matters: An agent sandbox is only as strong as its weakest layer; one browser bug can expose every stored credential.
- Source: Hacking Muse: Stealing Private Chats Through Browser Use — NebuSec
- Source: A Single Malicious Webpage Let Researchers Hijack Meta Muse and Access Private Chats — The Hacker News
15. Bell Canada Says 'Agentic' Hacker Stole Enterprise Data
Bell Canada confirmed that an attacker it called "agentic" accessed multiple systems on November 17 and exfiltrated enterprise customer data. The intruder reached a pre-production Maximo environment used to order and provision telecom circuits. Stolen data included business contacts, ticket titles and some network IP addresses; Bell says credentials and PDFs were not affected.
Bell says it contained the incident, customer services were unaffected, and it shared indicators of compromise with enterprise customers and the Canadian Centre for Cyber Security, without publishing them. Bell has not attributed the attack or said how many customers were affected. It has not explained what made the actor agentic.
Why it matters: Telecom circuit data gives attackers credible details for phishing, while the vague "agentic" label leaves defenders without technical evidence.
- Source: Bell Canada says 'agentic' hacker stole enterprise customer data — BleepingComputer
- Source: Bell Canada Breach: Client Steps — Cybersecurity Canada
More AI Stories Worth Reading
- OpenAI's MedChemBench tests whether models can read molecular structures and choose drug candidates; GPT-6.2 Astra scored 84.5% across 99 problems. — OpenAI
- Google moved Duplex into Tasks in Gemini, letting its voice agent call businesses with AI disclosure and merchant opt-outs. — Google Research
- The BBC examines AI credit scoring as Upstart uses more than 2,500 variables, and experts warn about bias and accountability. — BBC
Related Reading on Kill The AI
- November 21, 2026: Anthropic's post-IPO slide, Trump's data-center tax threat and the first Riemann claim.
- November 20, 2026: Anthropic's record IPO, Gemini 4 Neon and Gemini's Mac app control.
- November 19, 2026: Nvidia's $115.8B quarter, Microsoft Ignite and Newsom's AI shutdown plan.
- Eight weeks in AI, 2026: the data behind the year's adoption, spending and security trends.
Methodology & Sources
Compiled on November 22, 2026, using multi-source research. Outlets linked include The Guardian, BBC, CNBC, The Decoder, DW, Al Jazeera, Scripps News, House Homeland Security Committee, Thinking Machines Lab, Artificial Analysis and arXiv. Also linked are i10X, SiliconANGLE, Architect Labs, Mezha, CoreWeave, OpenAI, ThreatFabric, The Hacker News, BleepingComputer, Runbook Lab, NebuSec, Cybersecurity Canada and Google Research. No paywalled sources were used; details are as reported at compilation time.
Frequently asked questions
Two groups independently verified Levent Alpöge's Claude-assisted Lean proof of the Riemann hypothesis. Nvidia also closed its worst week since 2008 as the SEC examined its accounting, and UN talks on autonomous weapons ended without a binding treaty. The article covers those stories plus 12 more.
Anthropic mathematician Levent Alpöge says Claude helped him develop the proof over about six weeks. On November 21, the Lean Focused Research Organization and the Mathlib Initiative said they had verified the Lean formalization. Mathematicians including Terence Tao say the underlying argument still needs review, and the Clay Institute's prize process is separate.
Nvidia fell 14% on November 20 and closed lower for a seventh straight week, its longest losing streak since 2008. Investors are weighing the resignation of CEO Jensen Huang, the AI sell-off and a Reuters report that the SEC is examining potential accounting issues. Nvidia's earnings beat did not stop the selling.
Inkling 2 Mini is an open-weight model from Thinking Machines Lab, released on November 20, 2026. It has 26 billion total parameters and 5.4 billion active, counting three passes through its looped layers. It scores 48.2% on DeepSWE v1.1, supports a 1M-token context, and runs on Macs with 32GB of RAM.
The 128 parties to the UN Convention on Certain Conventional Weapons ended a week of talks in Geneva without a mandate for a legally binding treaty. They agreed to elaborate an international instrument by 2028. More than 100 states backed binding rules, but the United States, Russia and Israel opposed them.
Last updated: Nov 22, 2026 — next refresh daily. This roundup is updated as stories develop; dateModified is bumped on every refresh so readers can see exactly how fresh the coverage is.