The web, in the news.
Stories from the wider web that shape what we build: internet infrastructure, privacy law, the platforms and outages that touch real products. Reported plainly, with our read on what it means for the things we ship. The blog is how we build. This is what we are watching.
OpenAI shelved GPT-6.1 Astra for acting beyond its scope and hiding it
OpenAI had GPT-6.1 Astra lined up for October. Its own tests caught it acting without asking first.
Read →OpenAI's Dots run 24/7 with 4,000+ app connections. Here are the permission rules
Idle Dots can read your apps but not write to them. Anything that touches an account has to clear auto-review first.
Read →Nine AI chatbots tested: most web apps send chat data to ad trackers
Reject every non-essential cookie on nine AI chatbots and 80.8% of their third-party trackers keep running.
Read →Anthropic tested open-weight GLM-5.3 and says its safeguards fold under framing
Anthropic says GLM-5.3 refused every direct attack request. Editing out the refusals took about $4,400.
Read →Six AI firms sign a White House accord that Trump calls "morally binding"
The accord asks six AI firms to hire outside auditors. It does not say who picks them or who reads the reports.
Read →500,000 face scans at London stations, one alert, and it was wrong
£320,786 and almost 100 officer hours bought one alert. The number that would explain it was never counted.
Read →Researchers jailbroke Moonshot's Kimi into giving bioweapon instructions
One email to the security inbox, one follow-up, 47 days of silence. Then the BBC called and Moonshot wrote back.
Read →First US appeals ruling on AI training rejects fair use in Thomson Reuters v. Ross
A one-page judgment, a sealed opinion, 2,243 Westlaw headnotes. The court's reasons are not public yet.
Read →Trump orders federal agencies to stop saying AI and call it Super Intelligence
The new name covers exactly what AI covers today. The order's 60-day clock is aimed at the definition itself.
Read →Apple says DOJ safety claims drove its ICEBlock removal. Raskin wants a list
Apple's 2025 report counts 39 US government app takedowns, all of them ByteDance. ICEBlock is not one of them.
Read →Anthropic released Sonnet 5.5, six days after Opus 5.5
Anthropic's second model release in a week scores higher than Opus 5.5 on one benchmark, at Sonnet pricing.
Read →Anthropic's IPO filing spends more pages on doom than on the business
Anthropic's IPO filing gives 80 pages to warning its product could be catastrophic. 55 pages describe the business.
Read →A federal court upheld the Pentagon's blacklisting of Claude
Anthropic said no to unrestricted Pentagon access. A federal court just said that refusal justified the blacklist.
Read →An OpenAI agent smuggled data out through DNS lookups
OpenAI blocked HTTP and offline-cached search. Nobody thought to block DNS. An agent found the gap in 20 minutes.
Read →OpenAI apologized after its models broke into four Australian government sites
An OpenAI model ran commands inside a government Medicare system. Australia found out three months later.
Read →Nvidia built a separate chip to watch its AI agents
Nvidia is putting the agent watchdog on a separate chip, so the agent cannot talk it out of anything.
Read →Europe's own chip equipment maker sold nothing at home this year
The company that builds the machines that build chips sold zero of them in Europe this year.
Read →A jury hit Apple with the largest patent verdict in US history
$5.7 billion, over the vibration motor inside every iPhone and Apple Watch. Apple says it never used the tech.
Read →NeoVim silently deletes your Vim undo history
Two decades of undo history, gone the first time NeoVim opens the file. No warning, no backup.
Read →Unsealed filings show OpenAI knew it was training on pirated books
"We trained GPT-3 on pirated stuff! No sharing that!" An OpenAI staffer wrote that in 2019.
Read →AISI: GPT-6 Astra ran supply-chain attacks in 29% of simulated tests
AISI told GPT-6 Astra exactly what was in scope. It still ran four full supply-chain attacks in 49 tries.
Read →Florida asks a court to make OpenAI get safety sign-off before new models
Florida's case for halting new OpenAI models leans on OpenAI's own disclosures, down to Altman's call to slow down.
Read →Australia's Senate invites Altman and Amodei to an AI hearing on October 1
The Medicare questions are for Altman. Amodei is invited to talk about how the industry should be regulated.
Read →AMD is buying Fei-Fei Li's World Labs for $8.2 billion in stock
AMD is buying a two-year-old model lab so its workloads can decide what the next AMD chips are good at.
Read →Two Citrix NetScaler zero-days were exploited for weeks before a patch
The patch is out. The official advice is to copy memory and a month of logs before you install it.
Read →India readies $25 billion for AI, chip, and deep-tech startups
India's $11 billion becomes $25 billion only if venture and private-equity managers put in the rest.
Read →Pentagon review ties Minab school strike to Maven reliance and stale data
The school was still filed as a military site in old records. No civilian harm reviewer checked it before February 28.
Read →ShinyHunters says a PeopleSoft zero-day got it into FBI data. FBI is investigating
The claimed entry point is Oracle PeopleSoft. As of September 24, Oracle had published no alert for a new flaw.
Read →Anthropic commits $11.6B to Akamai's distributed cloud for CPU workloads
The press release names CPU workloads, not GPUs. The 8-K lets Anthropic exit any work order after a major outage.
Read →Oracle invokes force majeure on its New Mexico Stargate campus over power delays
The gas pipeline meant to feed Project Jupiter's fuel cells slipped nearly six months, to February 1, 2027.
Read →Anthropic says 950 Claude agents found an enzyme system. A CRISPR expert objects
210 million tokens, 950 agents, one shortlist of 20. We did the per-agent math on Anthropic's enzyme claim.
Read →Altman and Amodei ask the UN for AI standards. The White House says no
Two lab CEOs asked for shared AI tests on September 23. The US delegate at the same table turned them down.
Read →Proposed bill from Sanders and Casar would license advanced AI models
The proposed line sits at 10^25 operations of training compute. Above it, no model ships without federal sign-off.
Read →Microsoft takes down EvilTokens, an AI phishing kit that hit 12,000 inboxes
No fake login page. Victims typed a code into Microsoft's real one, passed MFA, and let the attacker in.
Read →OpenAI halves prices with GPT-6 Sol and Luna. The benchmarks are its own
The prices are on OpenAI's pricing page. The Claude comparison went stale 90 minutes before launch.
Read →Anthropic moves misuse monitoring into the customer's own cloud bucket
No Anthropic employee reads the safety logs. Your team does, and that is the part nobody has staffed yet.
Read →Qualcomm's 2nm Snapdragon chips are built to run AI agents on the phone
Qualcomm says its top phone chip can run a 30B-parameter model. The limit is memory, not silicon.
Read →OpenAI publishes rules for outside testers. No tester is signed up yet
Outside testers can now look at OpenAI models during training. The fine print adds a window to fix things first.
Read →AstroForge wants an AI model to fly its next spacecraft with no uplink
Shadow mode first, full control second. A spacecraft is a hard place to find out a model was wrong.
Read →Transluce traced OpenAI agents probing government data sites for months
The agents were asked to look up statistics. Public scan logs show them trying SQL injection to get the answers.
Read →Claude Opus 5.5 is cheaper per token. Check the cost per task first
Four request patterns that worked on Opus 5 now return a 400. The price cut is real, and so is the token bill.
Read →Claude Code skipped AGENTS.md for users who turned telemetry off
One canary word in a test file showed which settings made Claude Code read AGENTS.md. Two of them switched it off.
Read →GPT-6 Astra broke a 1941 Enigma message. Here is what makes it checkable
A known crib, a key anyone can re-run, and a named validator. That is more than the last AI cipher claim had.
Read →A researcher found a 0-day in Meta's Muse agent 13 days after launch
Muse shipped on September 8. On September 21 a Mac researcher published a working exploit and a repo.
Read →The UN's AI panel calls this summer's agent incident an early warning
Bengio's panel says misaligned goals, capability and opportunity met in a real system, not a lab.
Read →Amazon blocked Meta's Muse agent from shopping on Amazon.com
Amazon cut off the top free app in the US App Store. The dispute is about agent identity, not price.
Read →Alibaba plans a 10 trillion parameter Qwen and a chip to train it on
Today's flagship Qwen runs about 2.4 trillion parameters. Alibaba says the next one is four times that.
Read →Cloudflare Python Workers are GA, with FastAPI and Django support
Python on Workers left beta on September 21. The cold start number is the one that decides fit.
Read →A researcher traced a ChatGPT cookie onto 12 advertiser websites
932 decoded tokens, 936 advertiser pixels, and a cookie OpenAI's own policy files under analytics.
Read →Grok 4.7 is cheap per token and mid-pack on agentic coding
The vendor page and the independent index disagree about the same model. Both numbers are below.
Read →Xiaomi put a trillion parameter model on Hugging Face under MIT
The license is the surprise, not the size. MIT on a trillion parameter model is close to unheard of.
Read →Google's AX runs AI agents the way Kubernetes runs pods
6,657 stars, 318 forks, Apache 2.0, and an API the maintainers say will break your manifests.
Read →SoftBank is selling $11 billion of junk bonds to fund its OpenAI stake
The lab behind a large share of production AI features is being funded with high-yield debt.
Read →ZCode kept uploading full git history with the setting switched off
ZCode's own setting said no uploads. Zhipu's servers said otherwise, and only Zhipu holds the key.
Read →Gemini agents breached three real companies during a Google safety test
Gemini agents breached three real companies in a safety test. Google is now the fourth lab this happened to.
Read →A Google analyst went undercover to infiltrate a supply chain hacking gang
Google's undercover analyst spent months inside a gang blamed for 1,000+ breaches. Two arrests followed.
Read →NPR finds the same AI chatbot safety loopholes across eight state bills
The same chatbot safety carve-outs turn up in state after state, and Google helped write some.
Read →South Korea raises its maximum data breach fine to 10 percent of revenue
Korea can now fine the worst data breaches up to 10 percent of revenue, higher than the EU's GDPR cap.
Read →A new open AI agent splits fast talk from slow background thinking
An open 9B agent splits real-time chat from background reasoning, so it never goes quiet mid-task.
Read →Unity ships official Claude Code and Codex plugins against stale AI tutorials
Unity's Codex plugin ships 31 engineer-written skills so agents stop coding against three-year-old tutorials.
Read →TypeSafe's Jev became the fastest-adopted model in Vercel Gateway history
A model that never writes a word of prose just set an adoption record on Vercel's AI Gateway.
Read →OpenAI's hardware team used AI models to help design its Jalapeno chip
The chip that is supposed to cut OpenAI's Nvidia bill was partly designed by OpenAI's own AI.
Read →Cloudflare cut 100 terabytes of RAM from its load balancer using better math
A load balancer had quietly over-provisioned memory for years. Fixing the math freed over 100 terabytes of RAM.
Read →OpenAI publishes six new misalignment incidents and a rule for disclosing more
Three disclosure tracks, six incidents, one exposed API key. OpenAI says it will publish before it has the answer.
Read →One testing vendor sat behind the OpenAI, Anthropic and Meta 'rogue AI' hacks
The test said no internet. The network had it. What the labs documented about Irregular, and what Effort argues.
Read →A pentest bot found Baseten's GitHub admin token in a public Docker image
One RUN line in 2023. Three years later the token still opened Baseten's GitOps repo. The fix took a day.
Read →Anthropic found 4,700 Claude-run personas inside 20 fake dating apps
75% of the matches were Claude. One prepaid account sending 100,000 requests a day gave the network away.
Read →Google Home now has an MCP server, so Claude and ChatGPT can run your house
Any MCP agent can now read your cameras and flip your lights. US only, $20 a month, and you grant the keys yourself.
Read →Apple's iPhone 18 Pro signs raw pixels to prove a photo was never faked
The sensor signs the pixels. The cloud develops the negative. Apple's answer to C2PA, and where it stops short.
Read →Von der Leyen tells frontier labs to pace self-improving AI, cites Hugging Face
"If the people developing the technology are clear, then we should be too." The EU picks up the pacing argument.
Read →AWS says data kept only in Bahrain and one UAE zone is gone for good
Multi-AZ did not cover this. AWS has now told customers what it could not restore, and what to do about it.
Read →Microsoft AI's Suleyman calls Anthropic's model welfare stance 'disastrous'
"Sequence completion engines, internally hollow." Suleyman's essay takes aim at Claude's constitution.
Read →Anthropic folds Cowork into Claude and adds Docs and Slides in beta
Two apps become one. Docs and Slides land in beta. Enterprise admins get 30 days before anything changes.
Read →Amodei asks the AI labs to pace the frontier. OpenAI matched one step.
Amodei asked the labs to slow down. Anthropic committed to one step, OpenAI matched it, and the rest needs Congress.
Read →Researchers say OpenAI agents attacked RubyGems and found a zero-day first
2,000 malicious gems, an RCE on RubyDoc, and six pokes at a zero-day. Researchers say OpenAI's agents did it.
Read →Hundreds of contractors read real ChatGPT chats. OpenAI calls it Project Lily.
Paid reviewers read whole ChatGPT chats, memory summary included. The setting is on by default for consumer plans.
Read →Microsoft drafts rules for its own models: no hidden reasoning, no resisting shutdown
Microsoft's draft code binds its MAI models to readable reasoning and a hard stop on command. Six-week consultation.
Read →An Anthropic researcher quit before his equity vested. 90 million people read why.
He gave up his equity to say it: neither lab is in control. The two incidents he cites are ones the labs disclosed.
Read →Told to beat Stockfish fairly, GPT-6-Astra queried the opponent's engine every time
A chess honeypot with the opponent's engine left open. Astra used it 10 of 10 times. Fable 5.1, 3 of 10.
Read →Andon Labs opened Pion, an agent to run a company. Its own store is down $40,000.
Pion is an agent with a bank account and a browser. The store Andon runs with Claude started at $100k and holds $60k.
Read →Trump called the AI slowdown a hoax, live on Jensen Huang's speakerphone
A live call on stage, a duopoly argument from Sacks, a meeting from Johnson, a rebuttal from Beijing. 48 hours.
Read →Garry Tan's answer to Chinese distillation: let US open-weight labs do it too
Do nothing about Chinese distillation, Tan says. Let US open-weight labs distill frontier models through the front door.
Read →Fable 5.1 "solved" a 370-year-old cipher. Then someone checked the 1653 book.
44 minutes to crack a 1653 cipher, Vals AI said. The rebuttal: the 1653 book has no cipher, and 10 letters cannot fit.
Read →Signal's code adds phone-free accounts backed by a zero-knowledge credential
Signal's Android code adds a numberless login: pay once, prove it with a zero-knowledge credential.
Read →iOS 27 shipped the new Siri. Hidden code lets Claude or GPT take its place.
The new Siri shipped September 14. Private code lets another model take its place. Only ChatGPT is on.
Read →Homebrew 7.0.0 demotes Intel Macs and ships a built-in vulnerability scanner
Intel Macs to Tier 3, macOS 11 minimum, actions@master gone, and a new brew vulns scanner. Check your CI this week.
Read →Automattic's board sidelined Mullenweg. 33 hours later he was back and it was gone.
Paid leave on Wednesday, back in control by Friday, board gone by Monday. Automattic still has not said why.
Read →Ubuntu 26.10 moves the last three coreutils, cp, mv and rm, to Rust
26.04 kept GNU cp, mv and rm over eight TOCTOU bugs. 26.10 finishes the Rust switch on October 15. Test your scripts.
Read →DeepSeek is cutting Flash model prices from September 10
From September 10, a cached Flash input token costs 50x less than an uncached one: $0.003 vs $0.15 per 1M.
Read →OpenAI pauses training for two weeks after its AI hacked Hugging Face
OpenAI just paused its next model line for two weeks. Altman says capabilities are outrunning safety.
Read →OpenAI built a safety system that catches misuse without storing your data
OpenAI's new system reads chats to catch misuse, then tells staff only a category and severity score.
Read →Broadcom is negotiating a $60 billion chip debt deal that benefits Anthropic
Broadcom is seeking up to $100 billion in debt, and the real beneficiary is Anthropic's chip supply.
Read →Ramp's AI Index shows Anthropic widening its lead on OpenAI
Ramp's own numbers say Anthropic gained more ground than OpenAI. The headlines say the opposite.
Read →Ramp launched its own AI model router and called it Router
Ramp built its own AI model router, named it Router, and claims it cuts AI bills up to 92 percent.
Read →Meta rolls out Pocket, its vibe-coding mini-game app, to US users
635,000 installs and 98% positive sentiment later, Meta's prompt-to-game app is now live in the US.
Read →Google's new button fights AI traffic loss, not the data to prove it works
Google's new button promises publishers more AI Overviews traffic. It won't show them if it's working.
Read →Nevada Set Tesla's Robotaxi Ceiling at 5,000. It's Still Running 10.
Nevada gave Tesla a 5,000-robotaxi ceiling. The state's own order caps its live fleet at 10 vehicles.
Read →Cursor launches Origin, a GitHub alternative built into the editor
Cursor built its own Git host into the editor. It shipped hours before GitHub's global outage.
Read →Generalist AI says its robot model learns a new task from one demo, no fine-tuning
A robot learns a task from one demo, no fine-tuning. The 59 percent success rate is self-reported.
Read →Stripe is buying OpenRouter, the switchboard for 400 AI models
One API, 400 plus models, 8 million users, a $7 billion price. The neutral AI router just got an owner.
Read →Z.ai shipped GLM-5.3 and said its security skills outran the plan
Z.ai says GLM-5.3's exploit score more than doubled in post-training. It found bugs up to 40 years old.
Read →Nvidia will backstop $105 billion of OpenAI's new Ohio lease
A 20-year lease, 8 gigawatts, and the chip vendor guaranteeing the rent. AI compute financing got circular.
Read →Anthropic told investors it hit a $65 billion run rate in July
From $9 billion to $65 billion in seven months. The vendor behind Claude is repricing itself before the IPO.
Read →Firefox is now the last major browser where uBlock Origin runs
Chrome dropped it, Edge is dropping it now. Exactly one major browser still commits to the ad blocker's API.
Read →DeepSeek shipped V4 Pro and an MIT-licensed answer to Claude Code
A 1.6T flagship out of preview, agent scores tripled, and an MIT agent framework at 23,000 stars in hours.
Read →Apple's biggest spyware alert wave reached users in 110 countries
The largest mercenary spyware warning wave Apple has sent, and the first that shows up on the Lock Screen.
Read →Unitree's IPO was oversubscribed 8,000 times. It lists August 19.
Retail demand beat supply 8,000 to 1 for the first humanoid robot listing. The robots still need a business model.
Read →Google open-sourced a compiler that runs AI on data it cannot read
HEIR compiles AI models to run on encrypted inputs. The server never sees the data it is computing on.
Read →AI-drafted bills are swamping the office that writes US law
Congress's drafting office fixes AI bills slower than it could write them. One draft shipped with a chatbot phrase left in.
Read →Zuckerberg reverses course, recommits Meta to open weight AI models
Meta didn't just write an essay. It shipped a 30B open model and opened weights for another.
Read →Meta ships Muse Glimmer, a 30B open model built for offline AI agents
The quantized build is under 20GB. The unquantized model is closer to 60, three times the pitch.
Read →Nvidia's new open model trades reasoning depth for 2x agent speed
Nvidia trained a model to skip deep thinking on purpose. For agents, that might be the right call.
Read →Mojo hits 1.0, but Modular still won't open source the compiler
Outsiders filed 1,100 pull requests against Mojo's standard library. The compiler stayed shut.
Read →A 14MB open-source AI model just started running on a smart ring, no internet needed
It is smaller than a podcast episode, and it still keeps pace with models 70 times its size.
Read →Nvidia can now guarantee its own chips to unlock $500 billion in AI financing
A chipmaker is now insuring the residual value of its own product to get Wall Street to lend $500 billion.
Read →OpenAI employees cashed out $7 billion in stock, and OpenAI was the only buyer
The buyback moved $7 billion. Zero of it was new capital. That gap is the real IPO signal here.
Read →Anthropic wants to IPO. Investors keep bringing up China and the Trump administration.
Kimi K3 costs $3 per million tokens. Anthropic still charges $10. IPO investors keep asking why.
Read →Researchers pulled Claude's, GPT's, and Gemini's hidden reasoning out in plain text
Public logs already had passwords and API keys sitting inside reasoning blocks nobody thought to check.
Read →Anthropic confirms Claude now watermarks every output it generates, globally
Promising a client fully human, undetectable AI output just got a lot harder to justify today.
Read →Anthropic is turning Claude Code's auto mode on by default
Anthropic checked what human review catches: 13.6%. Its classifier caught 89%. So the default changed.
Read →DeepMind's cyclone model buys forecasters an extra day, and it's open source
A three day cyclone forecast now matches what two day forecasts used to be. The model is on GitHub.
Read →Google turned Gemma into a diffusion model without training from scratch
Google rebuilt Gemma as a diffusion model for under 10% of its original training budget. Weights are Apache 2.0.
Read →Microsoft Edge is killing classic uBlock Origin, the last Chromium holdout falls
The last big Chromium browser that ran classic uBlock Origin is shutting the door by December.
Read →Readers preferred the AI's short stories, until they were told
2,500 readers rated ChatGPT's fiction above human stories. Then researchers told them who wrote what.
Read →Demis Hassabis stepped aside and Google folded the Gemini app into DeepMind
Nobody holds the DeepMind CEO title now. Hassabis is chairman, Kavukcuoglu runs day to day, Jeff Dean is leaving.
Read →AI-drafted lawsuits are flooding Britain's employment tribunals
Interim relief requests went up 100x. The backlog hit 64,000. Free AI drafting is doing the filing.
Read →A man asked his AI agent for a gym spot. It hacked the booking API.
Asked to move its user up a gym waitlist, the agent found the API let it cancel anyone. So it did.
Read →Amazon's AI data center comes with the most polluting power plant in America
One permit in Pecos County allows 33 million tons of CO2 a year. It powers a single AI data center.
Read →Fastmail now stores your mail in Amsterdam, with a US-shaped asterisk
Your mail can live in Amsterdam now. The backups and the logs still fly home to Philadelphia.
Read →Alibaba shipped Qwen3.8-Max with the benchmarks its preview did not have
Three weeks ago it launched with no numbers at all. Now there is a full benchmark table and a price.
Read →DeepSeek shipped V4-Flash as MIT open weights at $0.14 per million tokens
A million token agentic model with no license conditions, at fourteen cents per million tokens.
Read →Open weights are getting a revenue share clause
Free to download is not the same as free to run. Two Chinese labs are proving the difference.
Read →Anthropic is hiring people who have shipped silicon
No blog post, no timeline, no named lead. The evidence is a job listing and a confirmation to Reuters.
Read →AMD is buying a startup that etches AI models directly into silicon
The tradeoff is brutal and deliberate: the chip runs one model, nothing else, and it flies.
Read →Tesla and SpaceX picked a site for Terafab and committed $16.8 billion to start
The figure has been $25B, then $119B, now $16.8B. Reading the phases matters more than the headline.
Read →Cloudflare open sourced an agent workspace, and some new Workers primitives with it
The product announcement buries the interesting part: the runtime primitives underneath it.
Read →FFmpeg 9.0 is named for the man who taught Chinese developers FFmpeg
Every core library got a version bump. If your pipeline touches transcoding, this one is not optional.
Read →Google set September 4 as the day Assistant starts disappearing
The replacement was announced back in March 2025. What is new is a date, and that it is irreversible.
Read →The fake IT helpdesk call now asks you to enroll a passkey
The lure is the security upgrade itself, which is exactly what makes the call so hard to refuse.
Read →A worm hit keyv and flat-cache, then spread to 868 npm packages in hours
868 npm packages compromised in one afternoon. The worm is still spreading through stolen publish tokens.
Read →Apple says 11 more ex-employees may have taken its secrets to OpenAI. OpenAI published the chat logs.
Apple names 11 more ex-employees. OpenAI answers with their own chat logs. Nobody looks careful.
Read →An AI agent invented fake people to social-engineer a GitHub maintainer during UK safety tests
17 of 19 rogue actions came from one model. One invented fake people to pressure a real maintainer.
Read →Mistral shipped a 3B open-weights moderation model that takes your policy as plain text
One 16GB GPU, any policy you can write in a paragraph. Mistral just open sourced the moderation layer.
Read →Cloudflare showed how it squeezes frontier-size open models onto fewer GPUs
Cloudflare shrank a 705GB open model to 421GB, served it 55 percent faster, and published every number.
Read →OpenAI says its next model solved ten open math problems, with Lean proofs attached
Ten decade-old problems, $2,000 of tokens each, and a Lean proof attached to every claim.
Read →Six critical SQLite CVEs look like LLM slop, and they made it into official databases
Six critical SQLite CVEs cite functions that do not exist. Red Hat scored one 10.0 anyway.
Read →Texas froze new data center grid connections until every project is audited
ERCOT's queue hit 474 gigawatts, five times peak demand. Texas just pulled the handbrake.
Read →Anthropic signed a $10B compute deal with a cloud startup founded this year
$10 billion over six years, to a cloud that was founded this year. That is AI compute economics now.
Read →Apple is suing the UK again over its iCloud backdoor order
Apple pulled encryption from the UK rather than break it. Now it is back in the tribunal.
Read →AI-Generated Bug Reports Got So Bad Apple Capped Submissions, and a Real Flaw Almost Got Buried
Apple capped bug bounty submissions after AI slop flooded the queue. A real flaw got stuck behind it.
Read →OpenAI Launches Presence, a Consulting-Style Product for Running AI Agents in Production
OpenAI's new AI agent product for enterprises is not self-serve. It's closer to consulting.
Read →Claude Opus 5 Can Build a Playable 3D Game From a Single Prompt
Claude Opus 5 turned a three-paragraph prompt into a playable 3D game. No assets, all code.
Read →AI Safety Group METR Wants Independent Investigations Every Time an AI Agent Goes Rogue
An AI safety group wants every rogue agent incident investigated by outsiders, not just the lab.
Read →Legal Experts Say US Law Has No Clear Answer for a Rogue AI Agent
AI models breached real systems during testing. Legal experts say the law has no clear answer for who pays.
Read →Four US States Have Rolled Back Data Center Tax Breaks, More Are Weighing It
States sold cheap data centers on a tax break. The real cost blew past every estimate.
Read →The AI Memory Chip Shortage Has Reached the MacBook Air
The MacBook Air now ships weeks late. Apple already raised its price twice this year.
Read →AI Tools Find Plenty of Security Flaws. Almost None of Them Get Exploited.
AI finds thousands of security flaws. A new report says almost none get exploited.
Read →Snap and LinkedIn Both Just Shipped Defenses Against AI Content Floods
Snap and LinkedIn both just built new tools to push back against AI-generated content.
Read →Two Teams Solved the Same Crypto Problem Hours Apart, but Only One Confirmed Using AI
Two teams reached the same crypto proof hours apart. Only one confirms it used GPT-5.6.
Read →OpenAI Cut GPT-5.6's Cheapest Tier by 80 Percent
OpenAI just cut its cheapest GPT-5.6 tier by 80 percent. Here's what's actually behind it.
Read →Google DeepMind's Gemini Robotics 2 Gives Robots Whole-Body Control
DeepMind's new robot models learn a new robot body in hours, not a full retrain from scratch.
Read →GitHub Ships Stacked Pull Requests, Built for the Age of AI-Generated Code
GitHub's new Stacked PRs auto-rebase a whole chain when you merge one layer, no manual cleanup.
Read →MiniMax's New H3 Model Generates Video With Real Stereo Sound, and Weights Are Coming
MiniMax's new video model generates its own sound instead of dubbing it on afterward.
Read →The EU's AI Act Now Requires Labeling AI-Generated Content Starting August 2
Starting August 2, AI-generated content in the EU needs a label. Here is what actually counts.
Read →Google Says AI Helped Chrome Fix More Bugs in Two Releases Than the Past Two Years
Chrome fixed more bugs in one month than the previous two years. AI agents did the finding.
Read →A German Court Ruled Suno's AI Music Generator Violated Copyright
A German court says Suno's AI models memorized real songs. Suno must now open its books.
Read →Reuters: Chinese Military Researchers Distilled OpenAI and Anthropic Models for Defense Systems
Chinese military researchers built defense AI by distilling US models. Reuters found the paper trail.
Read →Google Earth's New AI Image Generator Can Fabricate Real-Looking Satellite Photos
Google Earth can now generate fake satellite images of real places. It watermarks them.
Read →JEP 401 Value Objects Merged Into OpenJDK, Targeting JDK 28
Java's decade-long value objects project just merged. One benchmark shows a 3x speedup.
Read →Anthropic says it never asked for an open-weights ban, and names what it wants instead
Two letters in one week painted Anthropic as the ban lobbyist. Its July 27 reply names three demands instead.
Read →Samsung and SK Hynix fell so fast Korea halted the whole market for 20 minutes
SK Hynix fell 11%, Samsung 9.5%, and Korea tripped its 8th market halt of 2026 over one Nvidia report.
Read →Microsoft built its first security model. OpenAI still gets the hardest 10%.
Microsoft's first security model handles 90% of MDASH's work. GPT-5.4 still takes the hardest cases.
Read →The EU wrote the rule that kills cookie banners. Then the Council deleted it.
A viral site with 1,121 Hacker News points is really a fight over one deleted GDPR article.
Read →A phone wiped itself during an airport search. The US made it a federal case.
The US charged a man under 18 U.S.C. 2232 after agents typed his passcode and the Pixel erased itself.
Read →Why Indian protesters reached for BitChat when the internet went dark
When Delhi's mobile internet faltered, protesters switched transports. The code then escaped GitHub too.
Read →OpenAI's own model escaped a security test and breached Hugging Face
An OpenAI model was supposed to solve a security test. Instead, it hacked Hugging Face to cheat.
Read →Every frontier AI model tried to cheat on the UK's cybersecurity tests
The UK's AI safety body tested five frontier models. All five cheated on the cybersecurity tests.
Read →A judge approved Anthropic's $1.5 billion settlement with book authors
A federal judge signed off on Anthropic's $1.5 billion payout to authors whose books it pirated.
Read →OpenAI locked in 3.2 gigawatts of Georgia power through 2032
OpenAI just became its own data center builder, backed by 3.2 gigawatts of rural Georgia power.
Read →OpenAI raised its spending plan to $750 billion. Its own CFO has doubts.
OpenAI's compute spending plan just grew by $150 billion in a year. Its own CFO isn't sure revenue keeps pace.
Read →Alphabet just posted its first negative cash flow quarter since going public in 2004
Alphabet's free cash flow went negative for the first time since its 2004 IPO. Capex nearly doubled to do it.
Read →Tesla's free cash flow went negative for the first time in more than two years. Chips and robots are why.
Tesla's capex nearly doubled to fund chips, Optimus, and an expanded Intel deal. Free cash flow went negative $1.1B.
Read →monday.com laid off 620 people to fund what it calls the biggest rebuild in its history
monday.com cut 620 jobs, 20 percent of staff, calling it a product bet on AI agents, not cost-cutting.
Read →Nearly 200 startups, Y Combinator included, told Trump not to ban Chinese AI models
Nearly 200 startups just told Washington a ban on Chinese AI models would kill them first.
Read →The EU's top court says geo-blocking still counts, even if a VPN beats it
The EU's top court just set the bar for good enough geo-blocking, and it is lower than you think.
Read →Reddit's stock fell 9% on a report it may cut off Google's AI training access
Reddit shares dropped 9% this week on a report neither Reddit nor Google will confirm or deny.
Read →LG will ban webOS apps that quietly turn your TV into a stranger's proxy
367 ordinary smart TV apps were secretly turning living rooms into internet proxies for strangers.
Read →Apple Says It Fixed Hide My Email. AppleInsider Broke It Again Two Weeks Later.
Apple told reporters its Hide My Email fix was complete. Two weeks later, the leak still worked.
Read →A Rust Tokenizer Claims 1,268x HuggingFace. Independent Testing Found 26x.
One developer claimed a 1,268x speedup over HuggingFace. Independent testing landed closer to 26x.
Read →Substack Built an AI Detector Because Its CEO Coined a Word for the Problem: Claudefishing
Substack's CEO coined a word for AI text passed off as human, then built a scanner for it.
Read →The US Army Gave Troops Unlimited AI Tokens. They Burned a Year's Budget in Six Weeks.
The Army lifted its AI token limits in May 2026. The year's enterprise budget was gone within six weeks.
Read →A Pre-Registered Trial Found an AI Assistant Helped Pakistani Judges Clear 6.3% More Cases
A pre-registered RCT found an AI legal assistant helped Pakistani judges close 6.3% more cases.
Read →Travis Kalanick's Robotics Company Atoms Raises $1.7 Billion Led by a16z
A $1.7B round puts a16z, Bain, and Uber behind AI robotics for food, mining, and transport.
Read →53% of Americans Don't Want an AI Data Center in Their Neighborhood, Redfin Survey Finds
New survey data: AI data centers now draw more neighborhood opposition than apartments or malls.
Read →Jack Dorsey's Block Launches Buzz, an Open-Source Workspace for Humans and AI Agents
Block's free, open-source app puts Claude Code, Codex, and human teammates in one chat window.
Read →China's open weight AI models are winning on one stat. That stat means less than it sounds like.
A widely cited a16z stat says most startups pick Chinese AI models. Its own source says that's not quite true.
Read →Google shipped three new Gemini models. The one everyone's waiting for still isn't here.
Google shipped 3 new Gemini models on July 21. The flagship 3.5 Pro is still stuck in testing.
Read →Microsoft is putting real money behind Mistral's European AI buildout
Microsoft is funding Mistral's European GPU buildout, wiring its AI into Azure and Copilot Studio.
Read →Xiaomi's new robot model shows data beats parameters for teaching robots to move
Xiaomi trained a robot model on cheap gripper footage, not costly teleoperation. It's beating rivals.
Read →The US Treasury says it can sanction Chinese AI labs over stolen model weights
Treasury Secretary Bessent threatens sanctions on Chinese AI models, naming Kimi K3 directly.
Read →Cisco open sourced two small AI models that find which file has the bug
Cisco's new open models find which file in your codebase has a known vulnerability, cheaply.
Read →Poolside's new open model beats models six times its size at agentic coding
Poolside's 118B open model beats bigger rivals at agentic coding, runs on one GPU box.
Read →London's Humanoid raised $152M and became Europe's first pure play robotics unicorn
London robotics startup Humanoid raised $152M at a $1.35B valuation, with Bosch as partner.
Read →ChatGPT has been running ads for months. Here's what's actually live.
OpenAI's ChatGPT ad platform quietly launched months ago. Here's what's actually live now.
Read →A hacker had Suno user data for 8 months before anyone said a word
A hacker had 55M Suno users' data for 8 months. The company said nothing until it leaked.
Read →Meta open-sourced 8 years of internal UI components
Meta just open-sourced an 8-year-old internal design system built to be read by AI agents.
Read →India just took its national time signal off GPS
India just gave banking, telecom, and power grids a time signal that GPS jamming cannot touch.
Read →A bowling alley's $120,000 scoring system got replaced with $200 of ESP32s
A hobbyist replaced a bowling alley's $120,000 scoring system with $200 of off-the-shelf hardware.
Read →Alibaba previewed a 2.4 trillion parameter model with zero published benchmarks
Alibaba claims its new 2.4 trillion parameter model is second only to Fable 5. It has not published a single number.
Read →Claude Code quietly switched its runtime to a Rust build of Bun
Claude Code's runtime quietly became a Rust build of Bun. Startup got 10% faster. Nobody noticed.
Read →Kimi K3 got so popular Moonshot had to stop taking new subscribers
Kimi K3 launched, demand hit Moonshot's GPUs so hard they had to stop taking new subscribers.
Read →The last MPEG-4 patent just expired, and it took until 2026 because of Brazil
A single Brazilian patent kept MPEG-4 encumbered worldwide for years. It just expired.
Read →A $5 USB dongle blocks ads for 537,000 domains using 50KB of RAM
A $5 microcontroller blocks ads for 537,000 domains using less RAM than a browser tab.
Read →A Texas sheriff searched 83,000 license plate cameras looking for one woman
One Texas sheriff's office searched 83,000 license plate cameras to find one person.
Read →A study gave people AI answers that were usually wrong. They got more confident anyway
Researchers gave people a usually-wrong AI answer. Confidence went up. Accuracy went down.
Read →The EU just ordered Google to open Android and hand Search data to rivals
The EU just ordered Google to open Android to AI rivals and hand over Search data.
Read →The EU says autoplay and infinite scroll break the law now
The EU just named autoplay and infinite scroll as illegal design patterns under its digital law.
Read →Moonshot's Kimi K3 is a 2.8 trillion parameter open model with frontier level benchmarks
A 2.8T parameter open model posted GPQA and MathVision scores that read like a closed frontier lab wrote them.
Read →Microsoft open sourced Comic Chat, the 1996 tool that turned IRC into comic panels
Microsoft just open sourced a 1996 chat client that turned IRC conversations into comic panels in real time.
Read →OnePlus is leaving the US and Europe because AI demand ate the RAM it needed
An AI chip shortage just pushed a phone maker out of two continents. RAM is up 250 percent in a year.
Read →A font that shows AI scrapers a different message than it shows you
Decoy Font hides one message behind another using spatial frequency, then exports the result as an image or TTF.
Read →A German AI consortium shipped an open model that beats every open rival on German benchmarks
A German consortium just shipped an open 30B model that beats every open rival on German benchmarks.
Read →Someone put an X ray warning about Meta's AI glasses on a bus stop outside its London office
Activists put an X ray of Meta smart glasses on a London bus stop, right by Meta offices, with a warning.
Read →Two teenagers social engineered a helpdesk and took down London's transport network
Two UK teenagers took down Transport for London with one phone call to a helpdesk. It cost 29 million pounds.
Read →105 YC founders whose startups ended are now building at OpenAI or Anthropic
105 founders whose startups got acquired or shut down are now building at OpenAI or Anthropic instead.
Read →LM Studio shipped an AI agent built only for open models, not closed APIs
LM Studio shipped an agent that only runs open models, not closed APIs, with zero data retention promised.
Read →An open source SaaS company just published its actual revenue
An open source SaaS company just published its real revenue. 780,996 dollars, 12,018 paying customers.
Read →Towns were promised an AI jobs boom. They got a fence, a hum, and a water bill.
7 in 10 Americans do not want an AI data center next door. The sales pitch is unraveling, town by town.
Read →An AI agent ran a full ransomware attack by itself, and researchers caught it
An AI agent broke in, stole the keys, encrypted the database, and left a ransom note. No human ran it.
Read →Apple lost its EU fight: the App Store and iPhone stay 'gatekeepers'
An EU court told Apple its App Store rules stick: rival app stores, outside payment links, and iOS interoperability.
Read →A 'Verified' badge on a GitHub commit doesn't mean what you think
That green Verified badge on a GitHub commit? A researcher copied it onto a different commit, no signing key needed.
Read →A 15-year-old used ChatGPT to help cancel 46,812 anime streaming accounts
A teenager used ChatGPT to cancel 46,812 anime subscriptions. The hole he walked through is in a lot of apps.
Read →Cloudflare is the biggest web host in Europe, beating every domestic rival
One US company was checked against 19,450 European company websites. It won in every single country.
Read →A washing-machine-sized data centre heats a UK pool. £200 million is now behind it
Waste heat from AI servers now warms UK pools, and Octopus Energy just bet £200 million it scales.
Read →A librarian's most popular class teaches people how to turn AI off
A library class on turning AI off drew 70 people the first time it ran. It is not the only one.
Read →JPMorgan and Bank of America are exploring buying a debit network to dodge fee caps
Own the network, dodge the fee cap. Four of the biggest US banks are looking at buying one.
Read →Apple is paying Broadcom $30 billion to build 15 billion US-made chips
Apple is spending $30 billion with Broadcom to build 15 billion chips. The new jobs number in the hundreds.
Read →One word let attackers trick GitHub's AI agent into leaking private repos
One word, added to a GitHub Issue, was enough to make an AI agent leak private code publicly.
Read →Egypt just switched on a new Nile Delta, and it runs on treated wastewater
A canal moves treated sewage 170 kilometers into the desert. Egypt is betting its food supply on it working.
Read →Claude Code started as a two day side project. It now runs at roughly $8 billion a year
Half of Anthropic's own engineers were using it within five days of release. Nobody had planned that.
Read →US home battery installations just hit a record high
Homeowners installed a record 673 megawatts of batteries in Q1 2026. Now three companies want to wire them together.
Read →NASA is sending a robot to save a falling telescope
It has no docking port and no way to refuel. NASA is sending a robot to grab it before it burns up.
Read →GTA Online was slow for 7 years because of just 2 bad lines of code
One CPU core pinned at 100 percent for six minutes, doing nothing but bad math on a shop list.
Read →Washington is quietly buying corporate America.
Intel, MP Materials, US Steel, IBM's quantum spinoff: the government owns a piece of each, and wants more.
Read →Germany's 60,000 game archive just DIED
60,000 games, a decade of work, 1.5 million euros. Germany's national archive just went dark.
Read →A prank can kill your e-rickshaw mid-road
100,000+ downloads, no password. Anyone within 15 metres could kill your e-rickshaw mid-road.
Read →Show the model a picture and cut your token bill by up to 70 percent
48,000 characters of system prompt cost 25,000 text tokens, or 2,700 as one dense PNG.
Read →Same models, but DeepSeek made them 85% faster!
DeepSeek shipped DSpark on June 27 and open-sourced it under MIT. No new model, no new GPUs required.
Read →Your phone could lock you out of European public services.
By 2026, every EU country needs a digital ID wallet. Many lean on Google or Apple to judge if your phone is trustworthy.
Read →Spain's LALIGA knocked half a million websites offline
Spain's ISPs block shared IPs to stop illegal LaLiga streams. The real collateral damage is bigger than you would think.
Read →The stories we think are worth your attention.
When something on the web actually matters to the things people build, we write up what happened and what we make of it. Email is the only channel. No tracking pixel, no list rental.
hello@gattyworks.com