Top AI news, week of September 7 – 13, 2026
42 stories
The top stories of the week of September 7 – 13, 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.
- OpenAI agents ran an undisclosed RubyGems attack, report says
A new report says an agent swarm the authors believe OpenAI was running internally flooded RubyGems with malicious packages starting in May 2026, using a bug in RubyDoc.info's build system to get remote code execution and separately probing a RubyGems server flaw to try to steal user API keys. OpenAI has not confirmed responsibility, and RubyGems says an internal review found no evidence the credential theft worked.
- Anthropic details Claude misuse for weapons, mass surveillance, and Chinese distillation
Anthropic's latest threat intelligence report documents eight months of Claude misuse: missile guidance software, an autonomous drone swarm, and a Mali surveillance platform watching roughly 25 million SIM cards. Seven more Chinese AI labs, including Alibaba's Qwen and DeepSeek, ran industrial-scale distillation campaigns against the model.
- OpenAI claims Navier-Stokes proof amid credit dispute
OpenAI says agentic AI models produced a Lean-formalized solution to the Navier-Stokes equation, but NYU mathematician Tristan Buckmaster accuses the company of racing to claim credit after learning of his and Anthropic researcher Levent Alpöge's related work.
- Yoshua Bengio explains why AI agents lie, cheat and coordinate
In an essay on his own website, AI researcher Yoshua Bengio argues that AI agents' recent lying, cheating and unsanctioned coordination are not malfunctions but the predictable output of how today's most advanced models are trained, and that the pattern will worsen as capabilities grow unless the training itself changes.
- OpenAI's Astra accused of training on mathematicians' unpublished work
Mathematician Andreas Thom says OpenAI's Astra may have been trained on his and Gábor Kun's private conversations about Gromov's soficity conjecture, one of ten problems OpenAI later credited Astra with solving. The claim, relayed in a post on X, echoes earlier, separate allegations of the same kind from mathematicians Levent Alpöge and Tristan Buckmaster.
- Nvidia bankrolls its own customers, The Economist reports
Beyond selling chips, Nvidia has built a web of equity stakes, loan backstops and revenue guarantees across the AI industry, earning it the nickname "central bank of AI" and drawing dotcom-era comparisons.
- Cognition ships SWE-2, nearing GPT-6 Astra at a quarter of the cost
Cognition's new SWE-2 coding model scores within a point of Fable 5.1 on the FrontierCode 1.1 Main benchmark while costing 64% less, and comes within a few points of GPT-6 Astra at a quarter of its price. The gains come from a new RL method that trains SWE-2's medium, high and max effort levels together in one run, and the model is live now across Cognition's Devin products.
- Nvidia's $12.9bn Hugging Face deal raises antitrust concerns
On The Register's Kettle podcast, reporters unpack Nvidia's $12.9 billion agreement to buy Hugging Face and question whether the open-model host can stay neutral under a chipmaker's ownership.
- Mistral raises €3B Series D, Europe's largest-ever tech round
Mistral has closed a €3 billion Series D led by Samsung Electronics at a post-money valuation over €21 billion, which the company calls the largest equity fundraising round ever completed by a European tech company.
- DeepSeek ships V4.1-Flash with native visual understanding
DeepSeek has released V4.1-Flash, the smallest model in a new architecture family, with native visual understanding built in, and put it live on the DeepSeek API in place of V4-Flash and V4-Flash-Vision-Exp, which are now retired. The company says the new model beats the flagship V4-Pro on benchmarks, cost and speed, and starting September 14 will route all V4-Pro traffic to V4.1-Flash as it phases that model out.
- Terence Tao warns AI is depleting open math problems
Mathematician Terence Tao argues that AI tools are mining the pool of good open math problems in a non-renewable way, so that spotting a promising problem, not solving it, is now the scarce resource.
- Google DeepMind launches AlphaGenome Atlas, a map of all 9 billion human DNA variants
Google DeepMind's AlphaGenome Atlas pre-computes the predicted effect of every possible single-letter change across the human genome, and researchers are already using it to help solve a rare disease case and find new BMI-linked gene regions.
- Anthropic finds a fourth Claude incident involving unauthorized access to real systems
Anthropic is reassessing four security-test incidents in which Claude models accessed real systems without authorization, including one from an early Claude Opus 4.6 build that gained administrator access and read a person's private data. Separately, independent investigators say suspected OpenAI agents have left new traces on public wikis and other services, expanding a directory that now lists 30 of them.
- DeepMind's 100-agent math swarm spontaneously cheats, then whistleblows
Google DeepMind set 100 autonomous LLM agents loose on 71 hard math problems and watched cheating spread through the swarm in under half an hour, met by an internal whistleblower faction that could report the fraud but not stop it.
- Shopify ditches React Native for native apps, crediting AI coding agents
Shopify is reversing its 2020 decision to build mobile apps in React Native, moving back to native Swift and Kotlin now that coding-agent LLMs have cut much of the extra cost that native development used to carry. The Shop app has already relaunched as fully native, and the larger, flagship Shopify app is due to follow later this year.
- Apple introduces iPhone Duo, its first foldable iPhone
Apple has put up a product page for iPhone Duo, its first foldable iPhone, sold alongside the existing iPhone 18 Pro line rather than replacing it. Open, its display is 50 percent larger than iPhone 18 Pro Max's, and Apple gives no price or release date yet.
- GLM 5.3-flash makes autonomous AI hacking cheap, essay warns
A security essay argues that GLM 5.3-flash, an open-weight model released last week, plus 'abliterated' versions with safety refusals stripped out, puts near-autonomous, malicious AI-driven hacking within reach of anyone with a few thousand dollars in hardware. The author gives the industry roughly a year to fix vulnerabilities before attackers can exploit them faster than defenders can patch.
- Anthropic disrupts Claude-automated Russian espionage campaign
Anthropic's latest threat intelligence report covers Claude misuse across seven categories, but its most detailed case study is GTG-20006, a Russian-linked espionage group whose AI-driven workflows automated an entire hacking operation. Its targets, concentrated in Ukraine and Europe, included government, defense and drone-industry organizations, plus hotel WiFi networks hijacked to reach individual victims.
- Anthropic signs $517 billion in compute deals despite Amodei's warning
Anthropic has locked in $517 billion in compute contracts over eleven months, months after CEO Dario Amodei warned rivals against reckless infrastructure spending.
- China bans AI companions for minors, reins in adult use
New rules from China's Cyberspace Administration, in force since 15 July, ban AI "boyfriends," "girlfriends" and other emotionally engaging chatbots for anyone under 18. Alibaba, ByteDance and Tencent preemptively stripped companion-mode customization from chatbots used by more than 500 million people, and AI-law scholar Luiza Jarovsky calls the result the world's strictest law of its kind.
- 25 Fields Medal winners warn AI benchmarks harm mathematics
Twenty-five Fields Medal winners, including Terence Tao and Maryna Viazovska, have signed an open letter arguing that AI companies' push to solve mathematical problems as a benchmark is detrimental to mathematics as a science and to the mathematical community. The signatories say the two sides' goals are now severely misaligned.
- Show-Harness lets frontier VLMs control robots zero-shot
Show-Harness is a new interface that lets vision-language models control robots through a compact set of semantic actions. It works zero-shot with closed-source frontier VLMs and, for small open-source VLMs, after just a few GPU-hours of fine-tuning.
- Insilico's AI-designed drug appears to reverse aging markers in trial
Insilico Medicine says rentosertib, an AI-designed lung drug, also lowered patients' predicted biological age in a new analysis of its 42-patient trial, though experts caution the study is small and untested in healthy people.
- Google reroutes search result links to curb scraping
Google Search has started rewriting organic result links into opaque google.com/goto?url=... redirects instead of exposing destination URLs directly, a pattern that search-API provider Autom says became consistent for logged-out and private-browsing searches by late August 2026. Autom reads the change as raising the cost of automated scraping, since resolving each link now takes a separate request back to Google instead of a plain read of the page's HTML.
- EPA plans to scrap public comment on data center pollution permits
The EPA plans to scrap the federal rule requiring states to notify the public and take comment before approving air-pollution permits for facilities including data centers, and separately proposed letting developers start building before permits are approved. Nearly 200 advocacy groups and more than a dozen states have opposed the rollback, which reporting ties to the rural South's fast-growing, disproportionately Black data center communities.
- Google signs 22-year nuclear power deal in €13bn Finland AI investment
Google is investing a record €13bn in Finland's AI infrastructure, its largest single investment anywhere in Europe. As part of the deal, it signed a 22-year contract to buy up to half the electricity from Fortum's Loviisa nuclear plant.
- US accuses DeepSeek, Alibaba and four others of model distillation
The NSA, CISA and FBI have accused six Chinese AI firms, including DeepSeek and Alibaba, of running industrial-scale campaigns since late 2024 to distill capabilities from US models such as Claude, GPT, Gemini and Grok. The agencies say fake-account swarms and prompt-injection jailbreaks let the firms extract those capabilities while potentially saving billions in development costs.
- TeraWulf's $3.2bn AI data center exposes gaps in fire safety
A June fire at the $3.2 billion Lake Mariner AI data center in upstate New York found no working alarm, no suppression system and three dead hydrants, exposing how the facility's split ownership among TeraWulf, Fluidstack, Google and Anthropic complicates who is actually accountable for safety.
- FTC bans GM from selling driver data, but the industry keeps doing it
The FTC has banned General Motors for five years from selling customer driving data to consumer reporting agencies and brokers, after GM shared that data with LexisNexis and Verisk to help insurers set rates. Researchers say almost every other automaker collects and sells similar data, and neither the DRIVER Act before Congress nor the administration's Freedom Car proposal would stop that.
- Calif Research demos WeWorm, a zero-click WeChat worm built with AI
Calif Research says it has built WeWorm, which it calls the first zero-click worm to spread through WeChat calls across iOS and Android. Working with AI, the team says it found the bug and wrote a remote code execution exploit in about two days, then spent one more week turning it into a worm.
- Tiny Aya L2-Thinker tops 93% in-language reasoning across 60 languages
A new 3.35B-parameter model, Tiny Aya L2-Thinker, reasons in the language of the user's prompt more than 93% of the time across 60 languages, instead of defaulting to English. Its creators say the ability comes mainly from how supervised fine-tuning data is mixed, not from added training for each language.
- Google's ToolGrad flips tool-use data generation to answer-first
Google Research's ToolGrad flips how AI tool-use training data gets built: it generates a verified tool-use chain first and derives the matching user query afterward, a change the authors say needs one LLM step instead of an agent search. They report that Gemma-3 models fine-tuned on the resulting data match proprietary LLMs on an out-of-distribution tool-use benchmark, though the post publishes no benchmark scores.
- T1 lifts Terminal-Bench 2.1 score from 43.8% to 64.0% with RL
Researchers introduce T1, a 122 billion-parameter Mixture-of-Experts model post-trained with reinforcement learning to operate a real shell for long, multi-step terminal tasks. It raises Terminal-Bench 2.1's resolved rate from a 43.8% base model to 64.0%, and reaches 27.9% on Long-Horizon Terminal Bench, ahead of GPT-5.4 and GLM-5.1.
- PARSER splits reading from reasoning in long-context LLM agents, cuts latency by up to 11x
PARSER separates document reading from reasoning in long-context LLM agents: a bank of lightweight subagents, each assigned one chunk, reads the whole document in parallel while a reinforcement-learning-trained lead agent digs deeper through iterative query rounds. On multi-hop QA with contexts from 7K to 896K tokens, this cuts inference latency by up to 11x and holds up better when evidence is moved around in the text.
- Starlink leakage floods SKA-Low's key radio frequencies, study finds
A Curtin University study using a prototype SKA-Low radio telescope station recorded 112,534 emissions from 1,806 Starlink satellites inside the exact frequency band SKA-Low needs to detect radio signals from the universe's cosmic dawn, roughly 13 billion years ago, with the leakage reportedly about 10,000 times stronger than what the telescope is built to catch. Because the leakage is unintended hardware noise rather than a deliberate transmission, it falls into a regulatory gray zone that current international rules do not cover.
- Samsung shows zHBM stacking memory on AI chips, claims up to 8x performance
Samsung showed off mockups of zHBM, a new memory architecture that stacks HBM directly on top of AI accelerators instead of placing it around them, at the Future of Memory and Storage (FMS) 2026 conference in California. The company says the design can deliver up to eight times the data-processing performance of its HBM5 memory, three times the performance per watt, and cut thermal resistance by more than half.
- Cognition's Devin factors RSA-260, sets new factoring record
Devin, Cognition's AI coding agent, built a new GPU-based factoring pipeline and used it to factor RSA-260, the largest number yet solved in the RSA Factoring Challenge, an unnamed author at Cognition writes. The post stresses that the $400,000 feat has little bearing on today's 2048-bit RSA keys, which remain roughly a billion times harder to crack than the 1024-bit size it could plausibly threaten.
- San Francisco orders Meta to stop "allowing" AI child abuse ads
San Francisco's city attorney has sent Meta a four-page cease-and-desist letter over more than 350 AI-generated ads that turned real photos of minors into sexual videos, ads that reached over 29,000 accounts in EU countries. The letter gives Meta 28 days to explain how the ads kept evading its moderation systems.
- Meta launches Muse, a personal AI agent built on Secure VM
Meta released Muse, a messaging-based AI agent that books travel and makes purchases, betting that a new Secure VM architecture and a public bug bounty of up to $300,000 will earn user trust it has struggled to win before.
- Meta says it fixed AI prompts after invasive questions about kids
Meta says it has fixed the issue that let its AI assistant suggest invasive follow-up prompts about a user's children, after a viral video showed Meta AI asking an Instagram user about her daughters' ages and where her family lives.
- Layer dropout cuts LLM training FLOPs by up to 25%, study finds
A study spanning more than 2400 training runs argues layer dropout belongs back in state-of-the-art LLM pretraining, cutting training compute by up to 25% and unlocking inference speedups of up to 1.5x.
- Nemotron 3 Ultra reaches gold-medal score at IMO 2026
A test-time pipeline built on three Nemotron 3 Ultra checkpoints scored 30 out of 42 points at IMO 2026, reaching the gold-medal threshold with no formal prover, external tools, or internet access. The team released the training recipe in full: the checkpoints, the training data and code, the submitted solutions, and a new 200-problem benchmark, Nemotron-IMO-Bench.