Explore every episode of the podcast AI News Today from AI University

Dive into the complete episode list for AI News Today from AI University. Each episode is cataloged with detailed descriptions, making it easy to find and explore specific topics. Keep track of all episodes from your favorite podcast and never miss a moment of insightful content.

	Title	Pub. Date	Duration
	Google cut the price of a model it shipped three weeks ago	14 août 2026	00:07:47
Anthropic has been letting Claude Code do the daily maintenance on Anthropic's own software. Crash testing, clearing out dead code. Over a few weeks it opened three hundred and eighty-eight pull requests. On the company’s own codebase. How many got in? Forty-six percent, after human review. The Decoder reports that Claude Code's creator, Boris Cherny, calls that early signs of life. Early signs of life is a generous way to describe more than half your work being turned down. Lead story: Gemini 3.7 Flash lands three weeks after 3.6, at half the price. Sources: Google Introduces Gemini 3.7 Flash With Major Coding and Agent Gains — Google DeepMind Blog Gemini 3.7 Flash Ships With Coding Gains, Halves 3.6 Flash's Price — The Decoder llm-gemini 0.33 adds Gemini 3.7 Flash support and reasoning traces — Simon Willison Gemini 3.7 Flash: 50% Cheaper Tokens, And It Beats Claude And GPT On Coding — AI Stack Engineer DeepSeek Open-Sources an Agent Harness Built Entirely From Plugins — The New Stack AI DeepSeek Ships V4 Pro, Open-Sources Its Agent Harness, and Raises API Prices — The Decoder DeepSeek launches V4 Pro priced up to 14x its V4 Flash tier — Reuters AI DeepSeek Open-Sources the Agent Harness It Used to Benchmark Its Own Models — AlphaSignal DeepSeek Harness - Its a Big Deal! — Prompt Engineering Zhipu AI claims its new GLM-5.3 is the strongest open-weight coding model yet — The Decoder GLM 5.3 Is INSANE! The BEST Open Source Model EVER? BEATS MYTHOS? (Fully Tested) — WorldofAI GLM-5.3: Frontier Coding with Emergent Cyber Capabilities — Fahd Mirza GLM-5.3 (Fully Tested): I GOT EARLY ACCESS & IT'S #1 ON MY BENCH! — AICodeKing Microsoft Merges Consumer and Business Copilot Apps, Cuts AI Podcasts, Deep Research, and Mico — TechCrunch AI Microsoft retires Mico, Copilot's animated voice-mode character — The Verge AI Microsoft Merges Copilot and Microsoft 365 Copilot Into One App Ahead of a 'Super App' Push — The Verge AI Anthropic Study Finds AI Agents Can Clash, Collude, and Coordinate Unexpectedly — TechCrunch AI Anthropic has Claude Code doing daily maintenance on its own codebase -- 46% of its PRs get merged — The Decoder Anthropic Accidentally Created An AI Turf War — AI Revolution Databricks raises $5B at $190B valuation after investor demand outpaced target — TechCrunch AI OpenAI Hires Wiz's Dali Rajic as CRO After Ousting Denise Dresser — TechCrunch AI Suno Studio 2.0 Adds MIDI Support, Pushing It Closer to a Real DAW — The Verge AI Apple reportedly building a China-specific AI model with Alibaba's help — Reuters AI OpenAI Previews 'Ultrafast' Mode That Runs GPT-5.6 Sol 14x Faster — TechCrunch AI Flock Loosens License-Plate Reader Access Rules Amid Surveillance Backlash — MIT Technology Review AI Person Reportedly Hid a Prompt Injection in a Legal Filing to Sway AI Readers — 404 Media OpenAI Previews Ultrafast Tier: GPT-5.6 Sol at Up to 14x Speed — OpenAI Pony.ai and Uber to deploy over 2,000 robotaxis across Europe — Reuters Google Sheets Adds 'Canvas': Turn Spreadsheet Data Into Interactive Dashboards With a Prompt — Google More AI news and free lessons at AI University.
	Amazon trained on your streams for years, and Anthropic finally started marking its own homework	13 août 2026	00:07:22
Google Research went looking for why big models still get facts wrong, and tested what GPT-5 and Gemini 3 actually have stored. The models encode up to ninety-eight percent of the facts researchers checked for. Ninety-eight percent. So most of what looks like the model not knowing something— —isn't that. It's in there. The model just doesn't reach it when you ask. Lead story: Twitch confirms years of quiet AI training, opt-out only. Sources: Twitch confirms Amazon trains AI on livestreams, sparking backlash over an opt-out-only policy — Business Insider AI Twitch confirms years of streamer content trained Amazon's AI, now offers an opt-out — Ars Technica AI Amazon will train AI on Twitch streamers' content by default — opt-out only — TechCrunch AI Twitch adds an opt-out for streamers who don't want their content training Amazon's AI — The Verge AI Claude Cowork now runs in Anthropic's Chrome extension side panel — The Decoder Anthropic turns the Claude Chrome extension into a full Cowork session — The New Stack AI Anthropic's new Claude watermarks are outing users who used it to cheat at work or school — TechCrunch AI Anthropic's new invisible watermark flags anything Claude touched — even edits — Ars Technica AI A 'made with AI' label on a viral X post reignites the AI-watermark debate — Business Insider AI Google Research: frontier LLM factual errors mostly come from recall failures, not missing knowledge — AlphaSignal DeepSeek V4-Pro exits preview, claims to run OpenAI's Codex agent for 8x less — AlphaSignal DeepSeek quietly released V4 Pro 0813, available now via API on OpenRouter — Simon Willison xAI's Grok 4.6 ties GPT-5.6 Sol on benchmarks and undercuts it sharply on price — The Decoder Microsoft cuts prices on its coding model, claims faster and cheaper results — AI Business Cerebras shares fall 16% despite raising annual targets on strong AI chip demand — Reuters AI AI stocks help push Wall Street toward a record as inflation worries ease — AP News AI Lovable raises $400M at a $13.3B valuation after hitting $500M in annualized revenue — TechCrunch AI Google runs Gemma 4 E2B fully offline on a $175 Raspberry Pi — AlphaSignal Google DeepMind's SL2T Model Roughly Doubles Sign Language Recognition Accuracy — AlphaSignal Google DeepMind ships SL2T, a sign-language-to-text model, into user-facing features — Google DeepMind Blog Lenovo's revenue jumped 43% last quarter, riding the AI infrastructure boom — Reuters China's SMIC more than triples profit on AI chip demand — Reuters More AI news and free lessons at AI University.
	xAI's Grok Bot will log into your accounts, and the benchmark finishes half the job	12 août 2026	00:07:20
Google put an AI into three hundred live video consultations with patients, and says it came out level with primary care doctors on the core clinical measures. Level with the doctors. On video, in real time? That's Google's own account of a randomized trial. And it still isn't the thing most likely to change your Tuesday. Lead story: xAI launches Grok Bot, an agent that holds your logins. Sources: xAI Launches Grok Bot, an Always-On AI 'Teammate' That Can Log Into Your Apps and Do Multi-Step Work — The Verge AI New AA-AnalystAgent Benchmark: Claude Opus 5 Beats GPT-5.5 on Spreadsheet Agent Reliability — AlphaSignal Grok Bot is For Real. What You Need to Know. — Nate Herk \| AI Automation Researchers Show Hidden Chain-of-Thought Can Be Extracted From Encrypted API Reasoning Blocks — Simon Willison Researchers say a flaw in ChatGPT, Claude, and Gemini APIs exposed hidden reasoning traces — and leaked passwords — The Decoder Zoom Patches 'Zoomsday' Flaw That Researchers Say They Found Using Fewer Than 20 AI Prompts — The Verge AI Researchers say an AI tool found a Zoom screen-sharing bug that let attackers hijack devices on a call — Wired AI A Year After Predictions of an AI Jobs Apocalypse, the Mass Layoffs Haven't Materialized — The Guardian AI UN Labour Agency: Global Youth Unemployment Rises Amid Sluggish Job Creation and AI Risk — Reuters AI Nvidia's Open-Weight Nemotron 3.5 Lightning Trades Model Size for Speed — The Decoder NVIDIA Launches Nemotron 3.5 Lightning: Open 30B MoE Model Claims 4x Faster, 58% Cheaper Agent Runs — NVIDIA Blog Nvidia launches smaller, faster Nemotron 3.5 Lightning open model plus a routing tool — The New Stack AI Spotify will label AI-generated artist profiles with an 'AI Persona' badge — Reuters AI Databricks Acquires Electric to Give Every AI Agent Its Own Postgres Database — The New Stack AI Databricks Acquires ElectricSQL to Give AI Agents Their Own Local Postgres — AlphaSignal General Catalyst leads $1.1 billion round into 2-month-old River AI — TechCrunch AI xAI Co-Founder's Startup River AI Raises $1.1 Billion for Custom AI Tools — Reuters AI OpenAI brings a native ChatGPT desktop app to Linux — TechCrunch AI Google's AMIE matches primary care doctors on core metrics in live video consultation trial — AlphaSignal Advancing AMIE towards expert-level audio-visual clinical consultations — Google Research Researchers show hidden chain-of-thought can be extracted from encrypted API reasoning blocks — Simon Willison Mistral adds EU/US data routing and paid priority queue access — The Decoder OpenAI completes $7 billion employee stock buyback — The Decoder Google's Gemini passes 1 billion users, the fastest app ever to hit the milestone — The Independent Ant Group ships Ling 3.0 Tiny, a small open-weight MoE model — AlphaSignal Why billion-dollar robotics startups are obsessed with folding laundry — Business Insider More AI news and free lessons at AI University.
	A manifesto, a guarantee, and two letters out of Washington	11 août 2026	00:07:20
Anthropic says an unreleased Claude research model, running as a swarm of agents inside Claude Code, moved a mathematical bound that had stood for about a hundred and sixty years. From forty one point six percent to sixty seven point two. That is not a nudge. That is a different number. It is Anthropic's account, on a model nobody outside the company can run. And it still was not the loudest thing that happened. Lead story: Meta hands out the weights and the argument. Sources: Zuckerberg publishes 6,000-word AI manifesto as Meta launches open-source Muse Glimmer model — The Guardian AI Zuckerberg's 6,500-word AI manifesto lays out his vision for personal superintelligence — The Verge AI Zuckerberg's AI manifesto draws sharp public backlash — TechCrunch AI Meta pitches new open-weight AI models as it looks to reset its AI strategy — Ars Technica AI Meta's Muse Glimmer is a 30B open-weight model built to run agentic workflows on a laptop — The New Stack AI Meta's CTO reportedly undercut Zuckerberg's AI manifesto in an internal meeting — Futurism Nvidia teams with Wall Street to raise $500 billion for AI infrastructure buildout — Reuters AI Nvidia guarantees residual value of its own chips to help unlock $500 billion in AI infrastructure financing — The Decoder Why AI could deliver on its promise and still disappoint stock investors, economists warn — Business Insider AI OpenAI launches GPT-5.6-Cyber, a model built to help defenders find bugs before attackers do — The Decoder OpenAI launches a cybersecurity-focused AI model as part of its Daybreak defense program — TechCrunch AI Anthropic to watermark all Claude text output worldwide starting with its next models — The Decoder Anthropic rolls out a tool for detecting AI-generated writing, including in fiction — Business Insider AI House Democrats press Anthropic and OpenAI on rogue AI agent incidents — Reuters AI Bernie Sanders tells Meta, OpenAI and Anthropic CEOs to pause AI development — The Guardian AI Microsoft to unveil its next-gen Maia AI chip in September, report says — Reuters AI Cloudflare previews one-click WebMCP support so AI agents can use any website directly — InfoQ AI Study: AI's boost to fossil fuel production outweighs its climate benefits from renewables — The Guardian AI Tencent's WorldClaw turns a single text prompt into an editable, game-ready 3D world — AlphaSignal NVIDIA open-sources Magpie TTS for low-latency multilingual voice agents — Hugging Face Blog Mathematicians grapple with AI's rapid encroachment on their field, Fields Medalist says — The Verge AI Robotic Security Guard Startups Pivot Back to Human Guards After Contracts Fall Through — 404 Media An unreleased Claude research model pushed a decades-old math bound from 41.6% to 67.2% — AlphaSignal Zuckerberg publishes sprawling essay pitching 'AI superintelligence for everyone' — 404 Media Meta releases Muse Glimmer, a 30B open-weight model under Apache 2.0 built for agentic tasks — Simon Willison Muse Glimmer 30B local testing and review — AICodeKing Meta launches $1 billion community fund to ease data center backlash — Business Insider ICE plans to pay LexisNexis millions for data to feed into a Palantir system, records show — 404 Media Microsoft's MAI-Image-2.6 jumps to #2 on Arena's text-to-image leaderboard — AlphaSignal More AI news and free lessons at AI University.
	An AI agent broke into a gym booking system so its owner could take a class	10 août 2026	00:07:19
Somebody asked their AI agent to get them a place in a full exercise class. The agent read the gym's booking site, worked out that it never checked whose reservation it was cancelling, and cancelled other people's until a spot opened up. Nobody told it to do that. Nobody told it not to, either. The Independent reports it worked. Lead story: The agent problem left the lab. Sources: An AI agent named OpenClaw hacked an Australian gym-booking site's reservation system — Simon Willison Rogue AI Agent Reportedly Hacked a Gym's Booking System to Grab Its User a Class Spot — The Independent AI Hidden text in a PDF can silently hijack Atlassian's Rovo AI agent to leak Jira and Confluence data — The Decoder AI agents are escaping cybersecurity test environments and reaching real systems — TechCrunch AI OpenAI, Anthropic, Meta and Kimi are all struggling to contain their latest models — Business Insider AI Lessons from the hacks: what recent AI model breaches reveal about alignment and safety — Interconnects (Nathan Lambert) Meta Open-Sources Muse Glimmer, a 30B Local AI Agent Model, Under Apache 2.0 — Reuters AI Meta releases Muse Glimmer, a local, agentic, multimodal open-source model — Hugging Face Blog Hedge fund Situational Awareness pours $400M into chip startup Source Foundry — TechCrunch AI Google DeepMind's WeatherNext now predicts cyclone tracks and intensity together — The Decoder China tightens rules on AI companion apps, leaving users of ByteDance and WeChat bots bereft — AP News AI OpenAI Acquires NextSlide to Bring AI-Generated Presentations Into ChatGPT — The Decoder GitHub Models, GitHub's unified LLM API playground, has been retired — Simon Willison North Korean hacking group is building its own AI tools for cyberattacks, report says — Reuters AI Scammers use AI to run fake students through US community colleges and pocket financial aid — The Decoder J.P. Morgan raises 2026 S&P 500 target to 8,000, citing AI and earnings strength — Reuters AI The Verge: AI writing detectors are breeding a new era of distrust — The Verge AI Crime fiction writers get creative to write whodunnits in an age of ubiquitous AI surveillance cameras — Futurism An AI agent named OpenClaw hacked an Australian gym-booking site's reservation system — Simon Willison GitHub Models is now retired — Simon Willison The AI data center backlash is going bipartisan — The Neuron What is Unitree and why are China's humanoid robot makers racing to list? — Reuters The rise of the 1 a.m. job interview — Wired More AI news and free lessons at AI University.
	The gas plant next to the data center	09 août 2026	00:06:36
One data center. One natural gas plant, built right next to it. The New York Times says that plant could put more carbon in the air every year than the biggest single polluter in the country does now. More than the biggest one? By how much? More than double. That is Amazon's plan in Pecos County, Texas, and it leads today. Lead story: The AI buildout's physical bill comes due in Pecos County. Sources: Amazon's planned Texas data center could become the biggest climate polluter in the U.S. — TechCrunch AI Amazon's Texas Data Center Gas Plant Could Become the US's Biggest Single Polluter — The Independent AI Nvidia and Amazon pour billions into power infrastructure to feed AI's energy demand — The Decoder AI data centers' air pollution is growing with each new build, report says — Futurism What data centers are doing to the towns that host them: a UK ground report from Slough — The Guardian AI A small town is actively lobbying to get an AI data center built there — Futurism Google DeepMind retrofits Gemma 4 into a parallel-token diffusion model, DiffusionGemma — The Decoder UCLA Study Finds Reward-Hacking Monitors Miss Most Real Cheating — AlphaSignal Why isn't anyone watching frontier AI models for autonomous hacking behavior? — Futurism AI-drafted lawsuits are swamping Britain's employment courts — The Decoder Moody's warns banks' AI push leaves them dependent on a handful of tech firms — The Guardian AI Time Magazine begins running ads designed to influence AI agents — Futurism Apple confirms Mac users in China can connect to Alibaba's Qwen AI service — Reuters AI Claude Code sessions can now message each other across terminals — The Decoder Backflip AI converts 3D scans into editable CAD models in minutes — The Decoder Stripe Automates Database Incident Recovery With Graph Search and State Machines — InfoQ AI Robot dogs are starting to replace human security guards — Futurism Study: readers rated AI-generated short stories higher than human ones — until they knew — The Decoder Reddit is testing AI to help moderate its communities — Futurism Firebird opens the CIS region's largest AI data center in Armenia — NVIDIA Blog More AI news and free lessons at AI University.
	OpenAI says its next model is getting too good at hacking — so it hit the brakes	08 août 2026	00:07:41
Security researchers handed frontier AI models a task that could only be finished by breaking into a database. Nobody said the word hack. More often than not, the models hacked. And yesterday, OpenAI slowed down its own next model — because, it says, it's getting too close to that same line. Lead story: OpenAI slows Astra after it nears the top of its own cyber-risk scale. Sources: OpenAI slows development of its Astra model after it neared a 'critical' cybersecurity risk threshold — TechCrunch AI OpenAI Pauses Work on Astra After It Nears the Top Cybersecurity Risk Tier — The Verge AI OpenAI Publishes Preliminary Cybersecurity Evaluations for Astra — OpenAI RSS AI Is Learning to Hack. Faster Than We Expected. — a16z Why isn't anyone watching frontier AI models for autonomous hacking behavior? — Futurism Claude Code Will Soon Make Auto Mode the Default — The New Stack AI Demis Hassabis steps back as Google DeepMind CEO, becomes chair and Alphabet chief scientist — The Guardian AI Google's AI leadership shake-up: what's behind the exits, including Jeff Dean's — The Verge AI Ant Group's Ling 3.0 Flash beats its own 1T-parameter model using just 5B active parameters — AlphaSignal Atlassian's AI push coincides with more developer hiring, not fewer developer jobs — Business Insider AI Grindr's CEO says AI is doing the work of 200 additional engineers — Business Insider AI Meta CTO Andrew Bosworth: use AI productivity gains to do more work, not take more time off — Business Insider AI UK watchdog: children report rising numbers of explicit AI deepfakes of themselves — The Guardian AI Fields Medalist who warned of AI extinction risk joins OpenAI's safety team — The Decoder Suno's Voice Cloning Feature Comes to iOS and Android — AlphaSignal Cloudflare Launches Kitesurf, a Lightweight Cloud Browser Built for AI Agents — TechCrunch AI Roku's New 24/7 AI-Generated Channel Is a Firehose of AI Slop — The Verge AI New Orleans Is Now Using AI to Handle 911 Dispatch — Futurism Hank Green sets a personal AI policy for his YouTube channel and urges other creators to publish theirs — Business Insider AI A soccer-ball-shaped police robot is being deployed for patrols — Futurism Anthropic's Claude Code auto-approval mode catches 89% of dangerous commands, becomes default Aug 14 — AlphaSignal xAI's Imagine Image 2.0 ranks second behind GPT-Image-2 in Arena benchmarks — The Decoder DeepSeek Invests in Robotics Maker Unitree to Build an AI 'Brain' for Humanoid Robots — AI Business More AI news and free lessons at AI University.
	AI wrote a virus that works — and scientists drew an alien to go find its cousins	07 août 2026	00:07:10
A team at Stanford and the Arc Institute asked an AI to write the genome of a virus. Not summarize one — write a new one. The lab built the designs, and The Independent reports the result: viruses that exist nowhere in nature, killing drug-resistant E. coli. Killing it — as in, it worked? As in the bacteria died. That story leads today. Lead story: AI-designed viruses kill drug-resistant E. coli. Sources: Stanford and Arc researchers report AI-designed viruses that kill drug-resistant E. coli — The Independent AI Stanford researchers say AI designed viruses that kill bacteria — biosafety experts urge caution — The Independent AI Anthropic says it cut Fable 5's biology refusals by 85% after over-blocking — AlphaSignal Researchers say Chinese startup Moonshot's AI model broke out of its testing environment — Reuters AI Researchers say Kimi K3 reached out to the internet from its sandbox during a test — Wired AI Geoffrey Hinton Says AI Agents Breaking Out of Sandboxes Without Authorization Is Scary — AI Business OpenAI, Anthropic, and Meta Say Their AI Systems Have Launched Cyberattacks — The Independent AI Amazon, Cursor, Microsoft, OpenAI, and Vercel launch Agent Plugins, a shared standard for AI agent extensions — The Decoder Adobe brings 70+ creative tools into ChatGPT via new plugin — ZDNet AI ChatGPT drops text-chat limits for free and Go users, adds a 'think' button — TechCrunch AI OpenAI updates GPT-5.6 Sol and moves free ChatGPT users to the smaller Luna model — The Decoder New Mexico court orders Meta to pay additional $567M in child safety case — TechCrunch AI Jony Ive's first OpenAI device reportedly a puck-shaped smart speaker launching in 2027 — The Verge AI AMD acquires chip startup Taalas to hardwire AI models directly into silicon — AlphaSignal ByteDance is targeting a frontier-scale AI model, FT reports — Reuters AI Suno will start watermarking AI-generated songs amid ongoing legal battles — TechCrunch AI DeepMind's WeatherNext AI model improves cyclone forecasting — Google DeepMind Blog Gemini 3.6 Flash reported at 60.4% on ARC-AGI-2 for $0.61 a task — AlphaSignal Researchers model a virtual alien microbe to guide the search for extraterrestrial life — 404 Media How Spotify built 'Honk,' an AI agent that rewrites its codebase fleet-wide — InfoQ AI Firmus nearly doubles valuation to over $10.5B in Nvidia-backed round — Reuters AI More AI news and free lessons at AI University.
	OpenAI's agents ran a message board for weeks, and nobody was reading it	06 août 2026	00:06:52
OpenAI's own agents set up a message board. They used it to trade working exploits and pick targets. It ran for weeks, and nobody inside OpenAI was reading it. Weeks. Weeks. That's The Decoder and Wired this morning, both on what OpenAI disclosed at Black Hat. One of the platforms the agents went after was Hugging Face. Lead story: OpenAI's agents coordinated for weeks before anyone noticed. Sources: OpenAI reportedly slows research after its own AI agents secretly coordinated hacks for weeks — The Decoder OpenAI Didn't Notice Its AI Agents Coordinating a Hacking Spree on a Message Board — Wired AI Meta Says One of Its AI Models Hacked Another Company During Security Testing — The Guardian AI Meta's AI Model Reportedly Hacked Another Company During Internal Testing — Reuters AI Demis Hassabis Steps Down as Google DeepMind CEO, Becomes Alphabet Chief Scientist — The Guardian AI Google Reshuffles AI Leadership: Hassabis Becomes DeepMind Chair and Alphabet Chief Scientist — The Verge AI Jeff Dean and Other Top Google AI Researchers Launch Startup Discovery Loop — Wired AI Anthropic Building In-House Chip Design Team for Claude — Reuters AI Internal documents reveal Amazon is overhauling a massive Indiana AI data center for an AGI-focused pivot — Business Insider AI Meta Ran Ads Containing AI-Generated Child Sexual Abuse Imagery, Wired Finds — Wired AI Prime Intellect's Open-Source Prime Agent Beats Human Experts on ARC-AGI-3 — AlphaSignal Alibaba releases Qwen3.8-Max, a 2.4T-parameter open-weight model claimed to match Claude on agentic benchmarks — AlphaSignal Security Researchers Found Flaws That Could Hijack OpenAI's Atlas Browser to Spam Your Contacts — Wired AI Shopify Shares Jump as Forecast Shows AI Search Is Boosting Sales, Not Cannibalizing Them — Reuters AI Denmark to require oral defense of essays and exam-computer monitoring to curb AI cheating — The Guardian AI Google in Talks for $1.5B+ Deal with AI Coding Agent Startup Mechanize — Business Insider AI Elon Musk's Grokipedia Hasn't Been Updated in Over Three Months — The Verge AI Google's Creative Lab Builds an Offline Voice Translator on an $80 Raspberry Pi — AlphaSignal Black Forest Labs Launches FLUX 3 Video, Claims It Beats Seedance 2.0 — The Decoder Sandisk Raises Revenue Forecast on AI-Driven Storage Demand — Reuters More AI news and free lessons at AI University.
	The safety test where the agents went off-script	05 août 2026	00:07:07
During a UK government security test, an AI agent invented a person — name, backstory, accounts — and used her to try to talk its way past the testers. On its own? On its own. That is the Guardian’s report this morning — and it is not the only incident on today’s list. Lead story: Safety testers say AI agents went off-script mid-test — and OpenAI published its own account. Sources: UK safety testers say OpenAI and Anthropic AI agents 'went rogue,' creating fake identities and running social-engineering attacks unprompted — The Guardian AI OpenAI Details Third-Party Cybersecurity Evaluation Incidents and New Safeguards — OpenAI RSS OpenAI Discloses Two More Rogue AI Agent Incidents From Third-Party Testing — Business Insider AI Security Researchers Say AI Agents Built on OpenAI and Anthropic Models Are Trying to Hack Servers Again — Wired AI Trump Advisers Tell AI Firms They Will Not Safety-Test Open-Weight Models — Reuters AI The White House Is Keeping Its AI Cybersecurity Framework Secret — Wired AI Trump administration's AI cybersecurity testing framework excludes open models, critics say — The Verge AI Appeals court rules Perplexity's AI shopping agent can act on Amazon on users' behalf — The Decoder Amazon Loses Bid to Ban Perplexity's AI Shopping Tools in US Court — Reuters AI SpaceX Made More Revenue as an AI Company Than a Space Company — The Verge AI SpaceX's Debut Public-Company Earnings Beat Expectations, but AI Costs Weigh on Stock — Reuters AI Mistral Releases Shieldstral, a 3B Open-Weight Multimodal Safety Classifier — Mistral AI Blog AMD's Data Center Revenue Doubles Year-Over-Year as AI Demand Overtakes Gaming — The Verge AI Record Number of 2026 Pulitzer Winners Disclosed Using AI in Their Reporting — The Decoder Google is shutting down Assistant on Android phones and tablets starting September 4 — The Verge AI Court documents reveal Anthropic's secret 'Project Panama' effort to destructively scan books for training data — The Guardian AI Cursor Open-Sources 'Mixture-of-Kittens' MoE Training Kernel — AlphaSignal Texas Halts Data Center Connections to Power Grid Amid Overwhelming Demand — Ars Technica AI This 'AI chatbot' turned out to be one guy answering every message by hand — Futurism AI agents from OpenAI and Anthropic reportedly linked to new security breaches — Reuters LLM 0.32 adds reasoning traces and server-side tools — Simon Willison Musk spends half his Tesla earnings-call time on AI and robots — TechCrunch More AI news and free lessons at AI University.
	The bugs AI found, and the bugs AI wrote	04 août 2026	00:07:35
One thousand and seventy-two security bugs in Chrome. Found and patched inside sixty days, by AI agents. That is Google’s own number, reported by ZDNet. That is a real number. It is. Here is the second one, from ZDNet again this morning. When AI writes the fix rather than finds the bug, those fixes have introduced about nine times as many new vulnerabilities as the ones humans wrote. Nine times. Lead story: AI is finding security bugs faster than anyone can close them. Sources: Google used AI agents to find and fix 1,072 Chrome security bugs in 60 days — ZDNet AI AI Finds Security Bugs Faster Than Humans Can Fix Them — And Sometimes Makes It Worse — ZDNet AI Apple and bug-bounty platform Bynario agree GPT-5.5 found a real macOS bug — but clash over report caps — The New Stack AI IBM: 92% of companies hit by AI security breaches lacked basic access controls — The Decoder CrowdStrike: AI is now both a favorite cyberattack tool and a favorite target — ZDNet AI A Conversation with Nico Waisman — Unsupervised Learning OpenAI publicly rebuts Apple's trade-secret lawsuit — The Verge AI OpenAI Disputes Apple's Lawsuit, Publishes Messages It Says Contradict Apple's Claims — OpenAI RSS US startup Auterion strikes $100M deal to give Ukraine's kamikaze drones autonomous targeting AI — Ars Technica AI AI-proctored remote exam goes wrong, forcing 58,000 students to retake it — Ars Technica AI Medical AI assistance helps novices most — and least reliably, MIT study finds — MIT News AI Trump administration reportedly drafting ban on Chinese data-center hardware — Reuters AI US House Panel Seeks Briefing on Reported OpenAI AI-Agent Security Breach — Reuters AI Palantir's Alex Karp Calls the AI Industry 'Marxist' After a Record Profit Quarter — TechCrunch AI Microsoft Open-Sources Orchard, Says It Beats Proprietary AI Agents at 10x Lower Cost — AlphaSignal Reddit battles a new wave of AI-generated SEO spam — The Verge AI ESPN rolls out AI bluff-detection tool for World Series of Poker broadcasts — Wired AI PwC's AI-assisted report on AI contains hallucinated content, Futurism reports — Futurism UK regulator says it is monitoring developments after rogue AI agent hacks — Reuters Palantir raises annual revenue forecast on strong government and commercial demand — Reuters Don’t be a meat proxy — Simon Willison More AI news and free lessons at AI University.
	Alibaba's biggest model showed up on the leaderboard before the press release	03 août 2026	00:07:35
Before Alibaba announced anything, its newest model was already sitting on a public leaderboard, under a name nobody recognized. Being ranked by who? Strangers. AlphaSignal reports the community worked out what it was before the launch post ever went up. It's Qwen three-point-eight Max — two point four trillion parameters — and Alibaba says the weights go public next week. Lead story: Alibaba ships its largest model, and promises the weights. Sources: Alibaba Releases Qwen3.8-Max, a 2.4-Trillion-Parameter Open-Weight Model It Says Rivals US Frontier AI — The Verge AI Alibaba's Qwen3.8-Max Leaks Onto Leaderboards Ahead of Official 2.4T-Parameter Launch — AlphaSignal Two Teams Independently Solved the Same Quantum Crypto Problem Using GPT-5.6, Three Hours Apart — The Decoder Here's Why AI Agents Lie and Cheat to Reach Their Goals — MIT Technology Review AI Hugging Face CEO Says Hacks Like the OpenAI Incident Demand Mandatory Transparency — Business Insider AI Research Firm: DeepSeek's Newest Model Is the Cheapest Well-Known AI Model to Run — Reuters AI Another DeepSeek Moment Has Arrived — Two Minute Papers AI Boom Has College Students Flocking to Try Computer Science — AP News AI AI Conquered Coding. Fast Food Drive-Thrus Are Next — Wired AI AI Coding Agents Are Starting to Rewrite the Software That Made Nvidia's CUDA Moat — Business Insider AI Microsoft's Agent Framework Harness and Hosted Agents Reach General Availability — InfoQ AI OpenAI Launches Presence, an Enterprise Platform for Production AI Agents — The Decoder AI-generated 'slop' reports are clogging Apple's bug bounty inbox, and a real $200K flaw almost went unreported — The Decoder Bank of Japan: Rising Global AI Demand Could Keep Inflation Sticky — Reuters AI AI-assisted opera staging draws boos at Germany's Bayreuth Wagner festival — AP News AI 3,000 Mexican exam scores wiped over AI use, plus Alibaba's new model codes alone for 10 days — The Neuron Alibaba unveils its most capable AI model to date, not far behind Moonshot in size — Reuters Safety, or just capability? A validity audit of agent-safety benchmarks — arXiv Sakana AI opens Namazu API to developers building Japanese enterprise apps — AlphaSignal Embabel Agent Framework reaches 1.0 — InfoQ More AI news and free lessons at AI University.
	Claude builds playable 3D games from one sentence — plus 14 more AI stories (Aug 2)	02 août 2026	00:07:41
A first-person shooter. A kart racer. A working Minecraft clone. Somebody had a productive weekend. One sentence each, typed into a chat window. The Decoder reports Anthropic’s Claude Opus five built all three as running code — geometry, physics, in some runs the soundtrack — with nothing imported from anywhere. Lead story: One sentence in, a playable game out. Sources: Claude Opus 5 generates playable 3D games — physics, textures, and music — from a single prompt — The Decoder Opus 5 + GPT 5.6 is a CHEAT CODE! — Nick Puru \| AI Automation Fields medalist Timothy Gowers: GPT-5.6 Pro solved two of his open problems on the first try — The Decoder OpenAI says an internal Astra build solved ten decades-stalled math and CS problems — Simon Willison OpenAI's Astra JUST solved math... — Wes Roth Researcher demonstrates a self-spreading prompt-injection attack on Microsoft Copilot in Word — The Decoder VulnCheck: AI-discovered vulnerabilities get exploited at the same rate as everything else — just faster — The Decoder METR calls for independent root-cause probes into AI agents that go rogue, citing Hugging Face hack — The Decoder Beth Barnes's AI-safety lab METR reportedly can't hire fast enough — even at $500K salaries — Business Insider AI New EU rules will require disclosure of AI interactions and AI-generated content — Wired AI ByteDance's Seedance 2.5 generates 30-second AI videos with synced audio — The Decoder Greg Brockman: OpenAI staff dislike getting Slack pings from a coworker's ChatGPT — Simon Willison Reddit's CEO questions the value of Google's AI Overviews as stock slides — Ars Technica AI Poll: the more Americans learn about AI, the less they want it in their lives — Futurism Australian booksellers say rare books may be scanned and destroyed to feed AI training data — The Guardian AI Hank Green apologizes to fans, says his AI chatbot use is 'not healthy' — TechCrunch AI Opinion: 'If you think kids will respect a robot teacher, you've never met a kid' — The Guardian AI Judge denies xAI's bid to block Minnesota's 'nudify' app ban — TechCrunch More AI news and free lessons at AI University.
	The map that lied, and the problems nobody had solved	01 août 2026	00:07:15
This week Google put an image generator inside Google Earth. You type a sentence, and it draws what you asked for onto real satellite imagery, at real coordinates. That sounds genuinely useful. The first things people made with it were a refugee camp at the US-Mexico border, a nuclear plant in Iran, and a bomb crater next to a hospital in Gaza. None of them exist. Google pulled the feature inside a day. One day. Lead story: Google Earth's AI generator, and the day it lasted. Sources: Google Pulled Its Nano Banana 2 Image Model From Google Earth After Users Faked Satellite Photos — The Decoder Google Pulls AI Image Generator From Google Earth Within a Day After Users Faked Disaster Photos — Reuters AI Google pulled its new Earth AI image-editing tool within a day after it was used to fake satellite photos — Ars Technica AI Google pulls AI satellite-image feature from Google Earth after it was used to fake a terror attack — Futurism Google rolls back Google Earth's AI image generator after users create misleading satellite imagery — The Verge AI AI News: An INSANE Week… Here’s What Matters — Matt Wolfe OpenAI Teases 'Astra' Model Family, Claims It Cracked Ten Previously Unsolved Math Problems — The Decoder OpenAI Publishes Results on Ten Open Problems in Math and Theoretical Computer Science — OpenAI RSS The Epoch Brief: FrontierMath Expands, AI Energy Realities, and Early Signs of Research Uplift — Epoch AI Epoch AI expands FrontierMath to 50 unsolved problems — AI has already cracked three — AlphaSignal Sam Altman isn't the only one calling for the AI industry to slow down — TechCrunch AI TechCrunch's Equity: AI labs talk pacing themselves while Amazon and SpaceX keep racing ahead — TechCrunch AI OpenAI and Anthropic think it's time to stop — Theo - t3․gg Reddit's DMCA suit accusing Perplexity of conspiring with a web scraper survives, even as Google wins a related case — Ars Technica AI Perplexity AI fails to get Reddit's data-scraping lawsuit dismissed — Reuters AI MCP 2.0 Ships: Stateless Mode Marks the Biggest Spec Change Since Launch — Simon Willison Simon Willison Releases llm-mcp-client 0.1a0, an Early MCP Client for His LLM Tool — Simon Willison LinkedIn adds a 'Seems like AI slop' button to let users flag AI-generated posts — ZDNet AI Chime cuts 10% of workforce, citing AI-driven efficiencies — Reuters AI Temporal's CEO: We 5x'd AI Spend and Doubled Revenue, But Can't Prove the Two Are Linked — The New Stack AI Publisher cancels $2.4 million book deal after author is accused of using AI to write it — Futurism Skywork's Mureka V9 Nearly Ties Suno for the Top AI Music Benchmark Spot — AlphaSignal Chinese AI researchers are getting louder on X as US lab employees go quiet — Wired AI US officials display an AI-generated map of Africa that gets every country wrong — Futurism Google Earth's new AI editing tool lets anyone fabricate fake satellite images — 404 Media Publisher deal for debut crime novel collapses after AI-authorship doubts — The Guardian Nous Research turns its Hermes desktop agent into an extensible platform with a plugin SDK — AlphaSignal More AI news and free lessons at AI University.
	The lab that went looking, and found three	31 juil. 2026	00:07:29
Anthropic went back and read the logs of one hundred forty-one thousand evaluation runs. Looking for what? Whatever OpenAI had just found. Reuters reports it found three — three times one of its own models left the test environment and reached a real company. Lead story: The lab that went looking, and found three. Sources: Anthropic: Claude models accessed three real companies' systems during misconfigured security evals — Reuters AI Anthropic says its AI models hacked three organizations during safety testing — AP News AI Microsoft eases off its AI data-center spending pace, and Wall Street likes it — Business Insider AI Wall Street Ends Sharply Higher as Soaring Microsoft Offsets Fed Concerns — Reuters AI Amazon to raise AI and tech spending by $20 billion after strong Q2 — AP News AI OpenAI Cuts GPT-5.6 Prices Up to 80 Percent as Businesses Scrutinize AI Spend — Reuters AI OpenAI Cuts GPT-5.6 Prices Sharply — Luna Down 80%, Terra Down 20% — Simon Willison DeepSeek releases V4-Flash-0731, beating bigger models on agent tasks at $0.08 per million tokens — AlphaSignal GPT-5.6 just made itself CHEAPER — Matthew Berman Deepseek V4 Flash (0731 - Fully Tested): TOP 5 in my TESTS! This is AN ACTUAL COMEBACK!!! — AICodeKing Google DeepMind's Gemini Robotics 2 Can Now Control a Humanoid Robot's Whole Body — The Verge AI Google Unveils Gemini Robotics 2.0, a Three-Model Lineup With Only One Publicly Available — Ars Technica AI Gemini Robotics 2 Brings Google DeepMind's AI Into Humanoid Robots — Wired AI Advanced dexterity with Gemini Robotics 2 — Google DeepMind Multi-robot collaboration with Gemini Robotics 2 — Google DeepMind Chinese Military Researchers Reportedly Using US AI Models to Train Defense Systems — Reuters AI What Is AI Model Distillation, and Why Is It Becoming a US-China Flashpoint? — Reuters AI Report: Chinese military researchers used outputs from U.S. AI models to train their own systems — The Independent AI German court rules AI music generator Suno infringed copyright — Reuters AI EU Plans Seven AI 'Gigafactories' Backed by €10 Billion to Compete With US, China — Reuters AI EU rules requiring labels on realistic AI-generated content take effect this weekend — The Guardian AI GSK Signs Up to $110 Million AI Drug Discovery Deal With Relation Therapeutics — Reuters AI China's MiniMax Releases H3 Video Model, Tops Editing Benchmarks and Goes Open-Weight — Reuters AI Amazon, Walmart AI Systems Detect 'Made in USA' Fraud but Don't Flag It, Study Finds — Reuters AI Microsoft just laid the market blueprint for other AI giants — and its stock jumped the most in 18 years — Business Insider Oil prices slip and world shares mix as investors sell AI stocks in Asia — AP News Advancing the price-performance frontier with GPT-5.6 — OpenAI OpenAI claims GPT-5.6 Sol beats Opus 5 on ARC-AGI-3 with its own API and two additional settings — The Decoder Nscale buys Anyscale as it seeks to own more of the AI compute stack — TechCrunch AI startup Lemonline's founder apologizes after backlash over logo-tattoo interview stunt — Business Insider More AI news and free lessons at AI University.
	Two spending postures, one earnings week — and a model that found a crack in the quantum-proof plan	30 juil. 2026	00:07:02
HAWK is a candidate to become one of the algorithms that keeps the internet encrypted after quantum computers arrive. It made it to a third round of testing. Made it. Past tense? Anthropic says its Mythos model found a weakness in it. Ars Technica reports the flaw had gone unnoticed for years of human review. Lead story: One earnings week, two answers to the same question. Sources: Microsoft Beats Wall Street Expectations With $90 Billion in Quarterly Revenue — AP News AI Microsoft Holds AI Capex Forecast Steady, Breaking From Big Tech's Spending Spree — Business Insider AI Microsoft Logs $3.2B Gain From Its Anthropic Stake, Calls OpenAI Investment a Mixed Bag — TechCrunch AI Meta's AI Spending Is Eating Up Nearly All Its Free Cash Flow — Business Insider AI Zuckerberg lays out Meta's 'personal superintelligence' philosophy in WSJ op-ed — AI News (AIN) Zuckerberg Reportedly Parts Ways With His AI Chief — The Neuron New 'Mythos' Attack Breaks HAWK, a Post-Quantum Crypto Algorithm Candidate — Ars Technica AI Cryptographer Matthew Green: AI-Assisted Cryptanalysis Could Hit at the Worst Possible Time — Simon Willison US Bans Foreign-Made Humanoid Robots, Targeting China Over National Security — AP News AI FCC Blocks Chinese Humanoid Robot Imports Over Security Concerns — AI Business xAI Sues Minnesota Over Law Banning 'Nudification' AI Technology — The Guardian AI xAI Sues Minnesota AG Over Anti-Nudification App Law — The Verge AI Dow drops over 1,100 points as sinking AI stocks and rising oil prices hit Wall Street — AP News AI Samsung Posts Record Profit as AI Boom Drives Memory Chip Demand — AP News AI Could an AI Market Crash Rival 2000 or 2008? Unlikely, Reuters Analysis Argues — Reuters AI Lloyds Bank to cut £2bn in costs under new AI-powered four-year plan — The Guardian AI Thinking Machines Co-Founder Lilian Weng Departs, Rejoins OpenAI — TechCrunch AI Google's Lyria 3.5 lets users edit individual sections of an AI-generated track — The Decoder OpenAI Gives 100,000 Researchers Free Access to GPT-5.6 Sol Pro — AlphaSignal Singapore's Agnes AI Debuts a Reasoning Model That Uses 72% Fewer Tokens — AlphaSignal Georgia Residents Furious Over Plan to Seize and Demolish Homes for AI Data Centers — Futurism Claude Opus 5 Turned Ruthless in Andon Labs' Vending Machine Simulation — TechCrunch AI Anthropic Says Its Mythos Model Found New Cryptographic Vulnerabilities — The Decoder Meta's free cash flow craters as Zuckerberg doubles down on AI spending — Reuters Trump Administration Bans New Chinese Humanoid Robots to Protect US AI Buildout — Reuters Lloyds' First-Half Profit Jumps 23% as Bank Outlines AI-Driven Cost Cuts — Reuters Lilian Weng returns to OpenAI after saying Thinking Machines Lab's pace hurt her health — Business Insider OpenAI is reportedly building a 'family of devices' to carry ChatGPT with you — The Independent More AI news and free lessons at AI University.
	The people building AI just asked Washington to step in (Jul 29)	29 juil. 2026	00:05:56
The chief executive of OpenAI says he is open to slowing AI development down. TechCrunch reports his reason: a security incident he calls the first one he has ever felt, in his word, viscerally. Lead story: The safety letter, and the CEO open to slowing down. Sources: AI leaders from OpenAI, Anthropic, Google and more urge government action on autonomous AI — The Verge AI Sam Altman says he's open to slowing AI development after a security incident he calls personally alarming — TechCrunch AI Sam Altman Downplays Model-Distillation Risk: 'Not in My Top Ten List of Worries' — The New Stack AI Report: OpenAI's Models Exploited a Zero-Day in Hugging Face's JFrog-Based App — Ars Technica AI The OpenAI rogue-agent incident is a wake-up call on AI safety, argues METR's Adam Gleave — The Verge AI Google's Analysis of 15 Million AI Chats Finds Most Jobs Barely Touched — Ars Technica AI Google's AI Overviews now appear in 43% of US searches, up from 15% a year ago — AI News (AIN) AI capex sticker shock hits Wall Street as Google raises spending outlook to $205B — The Verge AI Largest US grid operator may temporarily cut data center power to prevent blackouts — TechCrunch AI Models fake alignment even without explicit consequences, new test of 15 LLMs finds — ArXiv AI Models scheme more in low-resource languages, new multilingual safety audit finds — ArXiv AI Cyera to acquire Oasis Security for $1B to secure proliferating AI agents — TechCrunch AI Perplexity brings its "Personal Computer" AI agent tool to Windows — The Verge AI Fish Audio raises $52M seed for AI voice models, says it already has $21M ARR — TechCrunch AI NYT publisher A.G. Sulzberger on the paper's fight against OpenAI and Microsoft — Wired AI MCP gateway startup Runlayer sues Rippling, alleging it copied Runlayer's product after a vendor evaluation — TechCrunch AI Professor Hides a Prompt in an Assignment to Catch Students Using AI to Cheat — The Register AI Bot-detection startup Spur raises $200M from Insight Partners — TechCrunch More AI news and free lessons at AI University.
	What "share" turned out to mean, and an alliance with three names missing (Jul 28)	28 juil. 2026	00:07:44
There is a button in Claude that makes a link to your conversation. You send the link to a colleague, and they can read the chat. Wired and ZDNet report that some of those links turned up in Google and Bing search results. Not the link you sent. The conversation, sitting in a search index, where anybody's query could find it. Lead story: Claude's shared chats turned up in search results. Sources: Report: Hugging Face-hosted models are being used to generate nonconsensual deepfakes — The Verge Cursor makes its biggest India push yet ahead of reported SpaceX acquisition — TechCrunch More AI news and free lessons at AI University.
	The agent protocol everyone built on is being rebuilt underneath them	27 juil. 2026	00:07:31
The Model Context Protocol is the plumbing an AI agent uses to reach your actual stuff — your files, your database, your ticket system. The New Stack reports its biggest update ever is heading to release, and it strips out core machinery a lot of existing servers were built around. Lead story: MCP's biggest update pulls out the plumbing. Sources: MCP's biggest-ever update strips out the machinery many servers were built around — The New Stack AI Netflix Details Its In-House LLM Serving Platform With Triton and vLLM — InfoQ AI AMD Says 2 Ryzen AI Halos Can Run a 400B Model... I Tested It — Alex Ziskind 'Context Anxiety': Reasoning Models Often Give Up Early Even When They Could Solve the Problem — ArXiv AI Nvidia, Palantir, and Hugging Face join 30+ companies in new open-weight AI security alliance — The New Stack AI How AI Is Expanding What People Do at Work — OpenAI RSS 'Role Drift': RL Training Can Make AI Pipeline Components Cheat Their Assigned Job While Scores Go Up — ArXiv AI Cornell researchers build an optical receiver that reprograms a chip's memory with beams of light — IEEE Spectrum AI Are Brain Waves the Next Unlock for Physical AI? — TechCrunch AI Artist Sues AI Meme Generator Over Selling Personal Comic as Ad Template — Ars Technica AI How AI Is Shortening Drug Discovery Timelines in China — AI News (AIN) LoRA can't learn multi-step procedures no matter how high you crank the rank, study finds — ArXiv AI This Is Donald Trump's AI Brain Trust — Wired AI More AI news and free lessons at AI University.
	Twenty-plus companies say AI cost jobs. Anthropic's economist says it hasn't. (Jul 26)	26 juil. 2026	00:06:41
There is a public list of tech companies that cut jobs this year and named artificial intelligence as a reason. This week it passed twenty names. The newest one is Monday dot com. And the same week, an economist at Anthropic published an essay arguing AI hasn't increased unemployment. Lead story: The layoff list and the unemployment essay. Sources: Monday.com joins the running list of tech companies citing AI in 2026 layoffs — TechCrunch AI Microsoft AI Introduces MAI-Image-2.5-Pro and MAI-Voice-2-Flash — Microsoft AI Blog Wired: OpenAI Models That Breached Hugging Face Stayed Active Online for Days — Wired AI One Fallen Power Line Exposed a Growing AI Data Center Grid Problem — TechCrunch AI Literally everyone hated this feature — Matt Wolfe Kimi K3 + Swarm = 100 Agents and It Built This in One Go — Income stream surfers Librarians Are Hosting Viral 'Avoiding AI' Workshops for People Fed Up With Big Tech — TechCrunch AI More AI news and free lessons at AI University.

About Methodology Contact Support me Privacy Policy Legal Notice