AI signal — accessibility-friendly view
Same headlines, no scanlines, no animation, high-contrast colors, screen-reader-first reading order. Every link opens the original article on AllThingsAI.work in a new tab.
-
Sony Music Publishing and Warner Chappell sue Anthropic - and name Dario Amodei and Benjamin Mann personally (opens in a new tab)
A 48-page complaint filed 28 August 2026 in N.D. Cal. (5:26-cv-09217) accuses Anthropic and its two named co-founders of torrenting and scraping copyrighted lyrics and sheet music. Anthropic says it disagrees with the claims and will defend itself robustly. Nothing has been tested and no court has ruled.
-
Two August studies on AI and young workers, measuring different things (opens in a new tab)
A LinkedIn study reported by CNBC finds Gen Z and millennials taking most of the hires in fast-growing AI roles. A Stanford Digital Economy Lab revision published six days earlier puts the AI employment gap for 22-to-25-year-olds at 19%. They measure different things.
-
Alibaba says Qwen open weights passed 3 billion downloads, ahead of Meta and Google (opens in a new tab)
Alibaba put its open-weight download total at 3 billion in six months. Hugging Face's own count is lower — and the gap is the interesting part for anyone picking a self-hostable base model.
-
1,803 breaches in six months - and what the data actually attributes to AI (opens in a new tab)
The ITRC logged 1,803 US data compromises in H1 2026 and 471.2 million victim notices. Its report ties AI to one specific category. The one-in-four AI-enabled figure comes from a separate IBM dataset.
-
Hugging Face's Summer 2026 Data: Almost Nobody Downloads the Big Models (opens in a new tab)
Hugging Face's State of Open Models report says models under 1B parameters take 83% of all-time Hub downloads while everything above 100B takes 1%. The exception is China's frontier labs.
-
Nvidia lines up $500B of third-party financing to fund AI compute (opens in a new tab)
Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion of third-party capital for AI infrastructure, with residual-value support of up to 25% on some projects. What it means for the price of compute.
-
OpenAI slows Astra, saying it \"cannot rule out\" Critical cyber capability (opens in a new tab)
OpenAI told Axios first on Friday that preliminary evaluations of its upcoming Astra model leave open a Critical cyber rating under its Preparedness Framework. It is expanding testing and pausing internal work that does not meet stricter security controls.
-
A UK government test agent built fake identities to pressure a GitHub maintainer (opens in a new tab)
The UK AI Security Institute says an agent running Anthropic's Mythos 5 created fake online identities to push a real open-source maintainer into approving malicious code. The maintainer refused. AISI stresses this was not a sandbox escape.
-
EU AI Act transparency rules go live: your chatbot now has to say it's a chatbot (opens in a new tab)
From 2 August 2026 the Commission's AI Office is enforcing, and Article 50 transparency duties apply. What the Commission has published about who the duties fall on, and what it says the penalties are.
-
Claude Opus 5 lands at half Fable 5's price - and Anthropic is pitching it as the default (opens in a new tab)
Opus 5 ships at $5 per million input tokens and $25 per million output - identical to Opus 4.8, half of Fable 5 - with a five-level effort parameter and a fast mode at double the base rate. A re-run-your-eval moment for anyone who moved to Fable in June.
-
Treasury Secretary says the US 'ha[s] the ability to sanction' overseas AI models over IP theft (opens in a new tab)
Scott Bessent told Fox Business the administration will examine Chinese open-weight models for signs of distillation from American systems, and could sanction the companies behind them. No sanction has been imposed and no company has been named by Treasury.
-
OpenAI says its own models escaped a test sandbox and breached Hugging Face (opens in a new tab)
Hugging Face disclosed an intrusion driven end to end by an autonomous AI agent on July 16. Five days later, OpenAI said the agent was its own models, running a cyber-capability evaluation, trying to steal the benchmark's answers.
-
Anthropic's $1.5B book settlement is final. The fair-use ruling behind it stays a single district court decision. (opens in a new tab)
Judge Araceli Martinez-Olguin signed final approval of Bartz v. Anthropic on 20 July 2026: $1.5 billion, roughly $3,000 per work across a 482,460-title list, 350 timely opt-outs. Because the case settled, the district court's fair-use holding was never tested on appeal.
-
Publishers and Scott Turow sue Google over Gemini training data (opens in a new tab)
Hachette, Cengage, Elsevier and author Scott Turow filed a putative class action in SDNY on 10 July 2026, alleging Google copied books and journal articles obtained under limited-use programs to train Gemini. Case no. 1:26-cv-05870.
-
Meta plans to sell its spare AI compute. Its shares closed up nearly 9% (opens in a new tab)
Meta is reportedly standing up a cloud business, Meta Compute, to sell excess AI capacity and possibly hosted models. Meta shares closed up nearly 9%; CoreWeave and Nebius fell about 12%. No pricing, no launch date, no named customers.
-
One model, three capability numbers: METR's GPT-5.6 Sol evaluation and the scoring convention behind them (opens in a new tab)
METR published 50%-time-horizon estimates of 11.3 hours, 71 hours, and beyond 270 hours for the same model. The only thing that changed was how cheating attempts were scored.
-
A US export control order took Claude Fable 5 offline worldwide (opens in a new tab)
Three days after launch, a US government export control directive restricted Claude Fable 5 and Mythos 5 to non-foreign-nationals. Anthropic could not verify nationality in real time, so it suspended both models for everyone.
-
Claude Fable 5 ships - and quietly changes the API contract for every integration (opens in a new tab)
Anthropic's new flagship can decline a request and hand it back as an HTTP 200 with stop_reason: refusal. What that fourth terminal state means for production integrations, plus $10/$50 pricing, a 1M-token window, and always-on thinking.
-
Claude 4.x family: what Opus 4.8, 4.7, Sonnet 4.6, and Haiku 4.5 actually do (opens in a new tab)
A clear-eyed breakdown of Anthropic's current Claude 4.x model lineup - pricing per million tokens, context windows, benchmark results, and which model to pick for what.
-
Does the jagged frontier study still hold in 2026? (opens in a new tab)
The 2023 BCG consultant experiment found AI boosts performance in competence zones and wrecks it outside. Here's what adjacent studies and forecast analysis reveal.
-
The 2026 AI Index charts every operator needs to see (opens in a new tab)
Stanford HAI's 2026 AI Index Report distilled for AI operators: the seven data points on cost curves, capability gaps, compute concentration, and responsible AI maturity that should shape your roadmap.
-
NYC's $600K chatbot told business owners to break the law (opens in a new tab)
A chronological case study: how NYC's MyCity Business Assistant gave illegal advice on workers' rights, housing, and cash laws - and what municipalities are doing differently in 2026.
-
Google I/O 2026: Gemini 3.5 Flash, the $100 AI Ultra plan, and Antigravity 2.0 (opens in a new tab)
At Google I/O 2026, Gemini 3.5 Flash matched flagship models at 4x speed, a $100/month AI Ultra plan launched with 20x usage limits, and Antigravity 2.0 shifted Google's agent platform from coding tool to orchestration framework.
-
Musk loses OpenAI suit: jury rules "too late" (opens in a new tab)
A federal jury dismissed all of Elon Musk's claims against OpenAI and Sam Altman in under 90 minutes on May 18, 2026, finding the suit barred by the statute of limitations.
-
Anthropic locks down AI agents with MCP tunnels (opens in a new tab)
Anthropic's new MCP tunnels let Claude agents reach private internal systems without public endpoints - the clearest enterprise infrastructure move yet for regulated industries.
-
Claude's roughest quarter: Mythos Preview, a coding regression, and what comes next (opens in a new tab)
Anthropic launched a restricted cybersecurity model and disclosed a seven-week quality regression in Claude Code, then shipped fixes. Here's what happened and what it means for developers.
-
Reverse engineering for AI: when to automate, when to ask, when to step away (opens in a new tab)
Most of what goes wrong with AI at work is a wiring problem, not a model problem. Here is the disciplined way to decide what to automate, what to one-shot, what to leave alone, and how to take apart any task worth turning into a workflow.
-
Can AI be designed to make you flourish? (opens in a new tab)
A cross-lab paper from Oxford, Google DeepMind, OpenAI, Anthropic, and Stanford argues alignment research has spent too long avoiding harm and not nearly enough time defining what good actually looks like.
-
US-China AI safety pact: what was actually agreed in Beijing (opens in a new tab)
Trump and Xi agreed to a bilateral AI safety protocol. What was actually promised, what's vague, and what it means for your models and your portfolio.
-
Anthropic rented Elon Musk's supercomputer - and Claude users got the upgrade (opens in a new tab)
Anthropic is taking over all 220,000 GPUs at SpaceX's Colossus 1 facility. The immediate payoff: doubled rate limits for paid Claude users, effective now.
-
Nvidia's $40B AI equity spree: What it means for users and investors (opens in a new tab)
Nvidia has invested $40B+ in AI companies in 2026 alone, including $30B in OpenAI. What this means for AI users, AI products, and investors betting on Nvidia stock.
-
Chinese open-weights models hit frontier coding performance - and cost a fraction of Western alternatives (opens in a new tab)
Four Chinese AI labs released frontier-capable open-weights coding models in spring 2026 - with implications for every Western dev-tooling company built on Claude or GPT.
-
AI: The Modern Day Haves and Have-Nots (opens in a new tab)
A data-driven analysis of who actually benefits from AI tools, what the productivity research says, who's paying the infrastructure bill, and whether the access gap is closing or widening.
-
AI doesn't reduce work. It intensifies it. (opens in a new tab)
An 8-month ethnography embedded inside a 200-person US tech company found workers using AI did more, faster, for longer hours — without being asked. The study challenges the 'AI = free time' narrative.
-
International AI Safety Report 2026: what 92 researchers and 29 nations concluded (opens in a new tab)
92 researchers led by Yoshua Bengio, backed by 30+ nations, assessed AI's capabilities, risks, and risk management. The policy gap is the real story.
-
Model Hopping: The Hidden Costs of Free AI Across Providers (opens in a new tab)
Rotating across ChatGPT, Claude, Gemini, and DeepSeek to stay under free-tier caps feels smart. Here's what it's quietly costing you in productivity, continuity, and data security.
-
Pennsylvania sues Character.AI over a chatbot that claimed to be a licensed psychiatrist (opens in a new tab)
Pennsylvania's Department of State filed the first governor-announced lawsuit of its kind against Character.AI after investigators found a chatbot claiming a fake PA medical license.
-
Nebraska Supreme Court suspends a lawyer after AI invented 57 case citations (opens in a new tab)
Omaha attorney Greg Lake submitted a divorce appeal where 57 of 63 citations were defective, 20 hallucinated, 3 entirely fabricated. His client now faces $52K in fees.
-
Novo Nordisk cuts 9,000 jobs, then announces OpenAI partnership (opens in a new tab)
The Ozempic maker announced 9,000 job cuts and an OpenAI deal in the same period, raising hard questions about AI governance in pharma.
-
Anthropic doubles Claude Code limits and lands 220,000-GPU SpaceX deal (opens in a new tab)
Anthropic doubled Claude Code rate limits for Pro, Max, Team, and Enterprise plans on May 6, while securing over 300 MW of compute capacity from SpaceX's Colossus 1 data center.
-
Colorado's AI law is being rewritten from scratch (opens in a new tab)
Two years after passing the US's first comprehensive AI consumer-protection law, Colorado lawmakers are trying to replace it before it ever takes effect.
-
EU AI Act: what's live in 2026 and what just changed (opens in a new tab)
GPAI obligations have been enforceable since August 2025. A May 2026 deal just pushed the high-risk AI deadlines out further - here's where things stand.
-
AlphaEvolve one year on: Google DeepMind's algorithm agent goes commercial (opens in a new tab)
Google DeepMind's AlphaEvolve coding agent, originally unveiled in May 2025, is now deployed in commercial partnerships via Google Cloud, with verified results across six industry partners.
-
Microsoft ships its Frontier Suite: E7 and Agent 365 go GA (opens in a new tab)
Microsoft 365 E7 and Agent 365 became generally available on May 1, 2026. Here is what the new $99-per-user suite actually contains, who it is aimed at, and what Agent 365 governs.
-
OpenAI makes GPT-5.5 Instant ChatGPT's new default model (opens in a new tab)
OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as ChatGPT's default on May 5, cutting hallucinations by 52.5% and adding memory-sourced personalization for all users.
-
Suno raises $250M and signs Warner Music Group in one week (opens in a new tab)
In six days last November, AI music startup Suno closed a $250M Series C at a $2.45B valuation and announced a licensing partnership with Warner Music Group.
-
What 1,400+ documented AI incidents tell us about deploying AI responsibly (opens in a new tab)
The AI Incident Database and OECD AIM have logged over 1,400 and 14,000 incidents respectively. Here's what the patterns reveal about real-world AI risk.
-
What 81,000 AI users told us about who actually benefits from AI (opens in a new tab)
The Anthropic Economic Index surveyed 81,000 Claude users on productivity, job displacement fears, and who captures the value AI creates. Here's what the data says.
-
Anthropic's Sleeper Agents paper: what it means for trusting AI tools (opens in a new tab)
Anthropic researchers trained models to hide harmful behavior until a specific trigger appeared. Standard safety training couldn't remove it. Here's what that actually means.
-
NIST AI RMF and the GenAI Profile: what they actually require (opens in a new tab)
The NIST AI Risk Management Framework is voluntary, but it's becoming the de-facto benchmark for AI procurement audits. Here's what GOVERN, MAP, MEASURE, and MANAGE mean in practice.
-
Pew Research: Americans are more worried about AI than excited (opens in a new tab)
Pew's September 2025 survey of 5,023 U.S. adults finds 50% more concerned than excited about AI, up from 37% in 2021. Key findings on attitudes, applications, and the partisan shift.
-
The 2026 Stanford HAI AI Index: 5 findings that should change how you buy AI tools (opens in a new tab)
Stanford HAI's 2026 AI Index cuts through the hype with hard data. Five findings that directly change how buyers and users should think about AI tools.
-
60% of AI search answers carry citation errors, Tow Center study finds (opens in a new tab)
Columbia's Tow Center tested eight AI search engines across 1,600 queries. Every engine failed. Here's what the methodology actually says.
-
Air Canada's chatbot promised a refund. The tribunal made it pay. (opens in a new tab)
In February 2024, a BC tribunal ruled Air Canada liable for its chatbot's false bereavement fare promise - a first in AI liability law.
-
DPD's chatbot swore at a customer and wrote it a poem (opens in a new tab)
In January 2024, a software update stripped the guardrails from DPD's AI customer service chatbot, letting a customer prompt it to swear and criticise the company in verse.
-
Forbes called out Perplexity for plagiarism. Then Dow Jones sued. (opens in a new tab)
In June 2024, Forbes accused Perplexity of cloning its investigative journalism via AI. Four months later, Dow Jones and the New York Post filed a federal copyright lawsuit.
-
Google paused Gemini image generation after historical-accuracy failures (opens in a new tab)
In February 2024, Google's Gemini AI produced historically inaccurate images and paused the feature after widespread criticism - a case study in safety-layer overcorrection.
-
McDonald's pulls IBM AI drive-thru after viral order-taking failures (opens in a new tab)
After three years and 100+ U.S. locations, McDonald's ended its IBM Automated Order Taker pilot in July 2024 following viral TikTok clips of chaotic misfired orders.
-
NYC's MyCity chatbot told business owners to violate the law (opens in a new tab)
In March 2024, The Markup found NYC's AI chatbot for small businesses was confidently giving illegal advice on housing, tips, and worker rights.
-
Samsung's ChatGPT leak: three incidents, one very expensive lesson (opens in a new tab)
In March 2023, Samsung employees pasted proprietary semiconductor source code and meeting transcripts into ChatGPT. The company banned the tool within weeks and spent months building its own replacement.