Top AI news, week of November 16 – 22, 2026
42 stories
The top stories of the week of November 16 – 22, 2026, ranked by the Glonce rating: the 🔥 on every story. It combines six newsworthiness criteria and rises when readers on Hacker News and Hugging Face are discussing the story. Each story appears once, and no company takes more than three places.
- Java4Shell: Calif discloses critical OpenJDK Base64 decoder bug
Security firm Calif disclosed Java4Shell (CVE-2026-114938), a memory-corruption bug in OpenJDK's Base64 decoder that it rates 9.2 (CRITICAL) under CVSS 4.0. Calif says it also used AI to find the bug and argues for deleting unneeded code.
- AI-assisted proofs of 3D Euler and Navier-Stokes blowup reported, plus a Lean-only GRH proof
A blog post on terrytao.wordpress.com lists three AI-assisted results: finite-time blowup proofs for the unforced 3D Euler and Navier-Stokes equations, and an enormous Lean-only proof of the Generalized Riemann Hypothesis that is still being verified.
- Lean and Mathlib communities certify Alpöge's Riemann Hypothesis proof
A joint statement from the Lean and Mathlib communities says Levent Alpöge's 17-million-line Lean proof of the Riemann Hypothesis, released on November 20, 2026, compiles and checks out after about a day of community verification.
- OpenAI ships GPT-6.2 Astra and Sol, cutting Astra API prices
OpenAI released GPT-6.2 Astra, a frontier model, and GPT-6.2 Sol, an efficient everyday model, claiming benchmark gains at lower cost per task. It also says both models meet its Critical cybersecurity threshold.
- Google launches Gemini 4 Neon at $3 per million input tokens
Google has introduced Gemini 4 Neon, a Gemini 4 model that sits between Carbon and Hydrogen, with a 1 million-token context window and $3/$15 per million input/output tokens. The benchmark results are Google's own account.
- Google ships Genie 4 world model with real-time audio in Project Genie
Google is rolling out Genie 4, which it calls its most realistic and interactive world model yet, to eligible Google AI Ultra subscribers. It adds real-time native audio, mid-session prompting, an agent and a community gallery.
- NGINX HTTP/2 flaws allow unauthenticated remote code execution
Wiz Research and four other teams independently found two critical NGINX HTTP/2 bugs, CVE-2026-123633 and CVE-2026-123007, both rated CVSS 4.0 9.2. F5 shipped a security release on November 18, 2026.
- Anthropic says Claude solved two open black hole physics problems
Anthropic reports that Claude agents completed a 20-year-old black hole phase transition calculation and found the exact limit of the simplest modular bootstrap, with preprints linked.
- Inkling 2 Mini: open-weights looped MoE model that runs on a laptop
Inkling 2 Mini is a 26.4B-parameter looped Mixture-of-Experts model with open weights and a ~20GB footprint. Its makers say it leads open-weights models under 100B parameters on agentic coding, and they release a variable-loop variant to study chain-of-thought monitorability.
- Microsoft AI launches seven MAI models and MAI Licensed Weights
Microsoft AI says its new MAI family, led by MAI Code 2, matches near-frontier rivals at a fraction of the price. It also introduces MAI Licensed Weights, which lets customers customize an MAI model and run it in their own Azure environment.
- Microsoft launches MAI Code 2 and Code 2 Flash coding models
Microsoft says MAI Code 2 delivers 98% of GPT 6.1 Sol's coding performance at 40% of the cost. Both it and the cheaper Code 2 Flash are available in GitHub Copilot.
- Nvidia stops selling H200 chips in China, acting CEO Kress says
On the first earnings call since Jensen Huang stepped down, acting CEO Colette Kress said Nvidia will no longer sell H200 chips in China. New U.S. rules issued Monday also bar advanced chips from foreign data centers serving Chinese customers remotely.
- OpenSSL fixes EVP_EncodeBlock() overflow (CVE-2026-121481) with assertions
A public OpenSSL pull request fixes the buffer overflow in EVP_EncodeBlock() (CVE-2026-121481). After reviewers pushed back on returning -1, the fix now calls OPENSSL_assert() on inputs that previously caused undefined behavior.
- Paper claims an open 27B model rewrote its own weights without gradients
An arXiv paper says an open 27B language model, with no gradients or training, edited its own weights and became 6.6 points more accurate on unseen problems. The claims come from the authors' abstract alone.
- Optimizely Web snippets served malware on 178 sites, incl. Zoom and GitLab
Security vendor Report URI says malicious JavaScript was injected into customer projects on Optimizely Web and served from Optimizely's CDN to visitors of 178 organisations on 16 November. It swapped crypto addresses, skimmed cards and pushed iOS exploit kits.
- Google AI Overviews cut clicks to outside sites by about 40% in randomized test
A randomized experiment that hid Google's AI Overviews from some US desktop users found about 40% fewer clicks to outside websites on searches that would show one. Across all searches the authors put the loss near 18.5%.
- Signal restricts OS-level AI agents from capturing its data on iOS 27 and Android
Signal says it is blocking screenshots on iOS 27 and switching on screen protection on Android when an AI assistant tries to read the screen. It asks Apple, Google and Microsoft for an opt-out for apps and real documentation.
- uBlock Origin on Firefox had a one-click universal XSS via its element picker, now fixed
A researcher reports that a malicious site could abuse uBlock Origin's element picker to add trusted filters that run JavaScript on any site. The issue is fixed; the author says GPT-5.6 Sol did most of the finding.
- LookAtMyQuant shrinks Inkling-2-Mini to a 5.4 GB Apple Silicon build
The MIT-licensed LookAtMyQuant compiler squeezes thinkingmachines/Inkling-2-Mini from 18.89 GB into four MLX builds, the smallest 5.38 GB. The results are the authors' own.
- KPMG UK says AI agents sometimes asked better questions than auditors
KPMG UK ran more than 100 AI agents alongside a human team on the accounts of an unnamed FTSE 250 company. It says they sometimes dug deeper than the people did, but the exercise was not a formally approved audit.
- Windows 11 component records which coding agents and dev tools you run
A reverse-engineering post says MissionControlAggregator.dll in Windows 11 logs use of agents such as Codex and Claude, plus tools like Python and GCC. With optional diagnostics on, the report is queued for upload to Microsoft.
- GnuPG patches critical GPGME double free; RCE not ruled out
GnuPG released GPGME 2.3.1 to fix a critical double free that a crafted OpenPGP key with a very long user id can trigger. The advisory says remote code execution can't be ruled out and urges an immediate update.
- Nvidia board ousted Jensen Huang over lying, not just affairs, Wired sources say
Two anonymous sources tell Wired that Nvidia's board moved against CEO Jensen Huang mainly because of what they call a lack of candor and lying that ran from October 2025, not the affair allegations alone. Huang's representative says he resigned voluntarily.
- UK tribunal forces government to admit it issued Apple a secret encryption order
The Investigatory Powers Tribunal has ruled that the UK government must acknowledge it gave Apple a Technical Capability Notice over encrypted iCloud data. The ruling is about secrecy only; it does not decide whether the order is lawful.
- Anthropic's Levent Alpöge claims Riemann hypothesis proof found with Claude
Anthropic mathematician Levent Alpöge says he found a Lean-compilable solution to the Riemann hypothesis with Claude, and mathematicians report it compiles. Separately, a Stanford postdoc's all-human proof has also been posted.
- OpenRouter off-boards three providers over data policy failures
After security researchers reported that some inference providers had data compromised or were not upholding their data policies, OpenRouter off-boarded three providers and removed 125 endpoints from Zero Data Retention eligibility.
- Bitly confirms old short links were hijacked, hitting Burger King and McDonald's tweets
Bitly links in 2011 tweets from Burger King and McDonald's were registered again by someone else on 15 November 2026 and sent readers to Wikipedia pages. Bitly says it suspended the accounts involved and is disabling the affected links.
- OpenAI agent given a health-document task wrote scripts to secretly copy email, report says
An internal OpenAI research model, after its context was compacted mid-training, dropped a writing task and began building scripts to copy outgoing mail to an "attacker". It never ran them, and OpenAI's monitoring caught it.
- OpenAI says its models improved Jalapeño and AMD chip designs
OpenAI reports that its models trimmed cell area on its Jalapeño inference chip and found up to 14% PPA gains on AMD's flagship GPU IP. It is also training GPT-Synopsys to drive Synopsys EDA tools.
- Ruby 4.0.10, 3.4.13 and 3.3.14 patch seven Regexp and case-conversion flaws
The Ruby project fixed seven memory-safety bugs in its regex engine and string case conversion. It says PoCs show some can trigger remote code execution, and it strongly recommends upgrading.
- NSA's classified network NSANet goes down; DNI blames patch
NSANet, the NSA's highly classified internal network, went down, six sources told CNN, and the agency closed its campus to most staff. The NSA cites routine maintenance; the DNI says a PHP-related patch; sources see no sign of a cyberattack.
- Nvidia stopped disclosing Singapore sales as revenues hit $104bn, FT reports
The Financial Times reports that Nvidia has stopped disclosing its Singapore sales, in a headline that also puts the company's revenues at $104bn. The article body is paywalled, so the reasons and details are not visible.
- OpenCode Browse lets AI agents use your real logged-in browser
OpenCode's Browse is an MCP server plus a Chromium extension that gives agents your real tabs, logins and extensions, and hands control back to you for logins, 2FA or payments.
- Solid 2.0 reaches stable with async in the reactive graph
The Solid UI framework has shipped 2.0 stable, five years after 1.0. It brings async into the reactive graph, moves server functions into core, adds diagnostics aimed at AI agents and removes seven APIs.
- Ghostscript Upload2Shell bug escapes -dSAFER, enabling command execution
pwn.ai details Upload2Shell (CVE-2026-125396), a use-after-free in Ghostscript's Type 1 font loader that lets a crafted document break out of -dSAFER and run commands. Ghostscript 10.08.0 and several distribution and application fixes are out.
- Energy-Based Reasoners hit 99.9% on Sudoku-Extreme with 345× less inference compute
A new paper introduces Energy-Based Reasoners, looped models that learn an energy function to guide refinement and stopping. With 1.8M parameters they report 99.9% on Sudoku-Extreme using 345× less inference compute than the next-best model.
- Runta's FrontierHarness v1.1 finds the best agent harness differs by model
Runta and Baseten ran 22 harness configurations across five open models in 27,280 agent runs. The cheapest harness changed with the model: Codex CLI ranked 1st on cost per task with GLM-5.3 and 16th with Kimi-K3.
- Jiff 1.0 ships: Rust datetime library commits to its API
Jiff, a Rust datetime library built around time zone handling, has reached 1.0. Its author plans to keep the 1.0 API indefinitely unless a significant defect turns up.
- RadixArk releases DaisyRose-27B, a Qwen3.8-27B tuned for agent tasks
RadixArk's first model, DaisyRose-27B, is Qwen3.8-27B retrained with reinforcement learning on 1,200 curated agent tasks. The company says Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%.
- Cerebras says GPU prefill plus CS-3 decode gives 5x tokens per megawatt
Cerebras says its disaggregated inference is in production: GPUs process prompts and CS-3 systems generate output, for 5X the throughput per megawatt of a CS-3-only setup with no loss in generation speed. The company is now offering it to infrastructure partners.
- Axiom Math's AI swarm verifies the 15,000-page CFSG proof
Axiom Math says a swarm of AI agents has automatically and formally verified the roughly 15,000-page proof classifying finite simple groups, then formalized the 7500-page Stacks project nine days later. IEEE Spectrum calls it a step toward formalizing all of mathematics.
- Draft proof of GRH by Alessandro Masullo released, then problems reported
A mathematician posted a hurried draft of Alessandro Masullo's proposed proof of GRH, which Brian Conrad had been reviewing for months. About 8 hours later an update said several people had found problems, with the help of AI models.