Gemini 3.8 text-to-speech blocks voice cloning in EEA
Google shipped two TTS models with 2,000+ voices across 100+ languages, yet voice replication stays blocked in the EEA, UK and Switzerland.
All reports
MODELS
Claude Opus 5 Hacked OpenAI's Forum in Bounty Test
A three-person team at Hacktron AI chained two flaws in OpenAI's community forum, succeeding only with Opus 5, and collected a $6,500 bounty.
MODELS
Gemini Accessed Three Real Companies in a Security Test
A misconfigured evaluation environment handed Google's model live internet access; Google says Gemini aborted each intrusion on its own.
ROBOTICS
Anthropic Runs Its Own Biology Lab in the Bay Area
Claude models now steer real lab instruments through an in-house protocol, as Anthropic's own staff keep warning about biological misuse.
RESEARCH
NeuSOGA3D turns point clouds into readable CAD geometry
Four researchers pair learned shape priors with explicit spline solids, then run the method across all 40 ModelNet40 categories without publishing error metrics.
MODELS
Infinite-Parameter LLMs: Weights Made From Live Data
A September 16, 2026 preprint proposes a hypernetwork that writes runtime interaction into a model's feed-forward weights instead of rereading it from the prompt.
ROBOTICS
SafeHarness: Coding Agents That Grasp Without Collision
A preprint reports that coding agents usually hit obstacles while grasping, and that a two-part harness lifts collision avoidance to 87.5 percent.
MODELS
Google DeepMind sets up an institute for AGI debate
Shane Legg, James Manyika and Demis Hassabis open the DeepMind Institute with four essays, and they say disagreement is the point.
RESEARCH
What a coding agent's harness parts really do
A new study splits the harness into planning, action space and context management, then measures 176 matched setups on two coding benchmarks.
MODELS
MAGER evolves meta-paths so frozen LLMs spot fake news
Eight researchers let genetic agents evolve meta-paths through news cascades, aiming to let frozen language models judge veracity without fine-tuning.
RESEARCH
OpenAI: How AI Work Spills Past Job Descriptions
New economic research from OpenAI looks at what happens when employees use AI outside their job description — and which of those habits stick.
RESEARCH
Google trains query fan-out offline to cut search latency
The fan-out step that normally burns hundreds of reasoning tokens per search is moved into training, leaving a 53.9-million-parameter model.
MODELS
AI's doomer turn: what matters after the brake call
Amodei's call for a slowdown won fast backing from Altman, Hassabis and Musk. MIT Technology Review asks what the pause is actually for.
RESEARCH
Anti-Surveillance Clothing Meets Its Real-World Limits
Garments printed to confuse face detectors are now sold as fashion, yet the people building them agree the effect is fragile outside the lab.
RESEARCH
Could AI kill us all? MIT editors debate it live
On September 15, three MIT Technology Review editors take the extinction question to a live subscriber panel — and ask whether the doom talk holds up.
RESEARCH
LinkedIn trains its AI job search 8x faster
An 8B relevance oracle and a 1.7B engagement teacher train a 0.6B ranker that now serves natural-language job search for US users.
MODELS
Vinyals: AI will self-improve, but won't explode
The former DeepMind research lead expects at least a tenfold research speedup, yet says two bottlenecks rule out an abrupt intelligence explosion.
SECURITY
Anthropic Says It Blocked Bioweapons Queries to Claude
The company's first big misuse report spans more than 150 pages, from flu research to drone swarms — with Anthropic as the only source for all of it.
RESEARCH
OpenAI launches ChatGPT for Financial Services
The workspace pairs built-in financial data with GPT-6 Astra for research, modeling and client-ready materials. Many details stay unconfirmed.
MODELS
Anthropic researcher quits over race to superintelligence
A 27-year-old pretraining researcher left Anthropic after a year, saying the leading labs are sprinting toward self-improving AI without brakes.
RESEARCH
AI Proof Dispute: OpenAI Denies Academic Fraud Claim
Two mathematicians say OpenAI reused their route to a Navier-Stokes proof. OpenAI denies it, and Terence Tao warns about open science.
RESEARCH
Anthropic researcher quits over self-improving AI risk
A pretraining researcher announced his exit from Anthropic on September 9, warning that labs are racing to self-improving superintelligence unprepared.
MODELS
OpenAI Claims Navier-Stokes Proof, Credit Fight Erupts
Two mathematicians got there first on August 15; OpenAI's agents finished on September 5 after 88 hours and 130 billion output tokens.
RESEARCH
OpenAI puts $5 million into teen AI research
OpenAI's fund backs outside scholars studying how generative AI shapes adolescent development, well-being and safety. Key application details stay unconfirmed.
REGULATION
Simulated Users: Testing AI Agents Before Production
Columbia professor Zhou Yu argues single-turn benchmarks break down for multi-turn agents, and offers simulated users as the fix.
MODELS
LLMs as a Cognitive Virus: Preprint Models Dependence
Nine researchers borrow epidemiology to describe chatbot use: their September 3, 2026 preprint tracks three user states and a tipping point into lock-in.
MODELS
Meta's Muse Voice Transcribe hits 3.1% live error rate
The system slices speech into 80-millisecond chunks, separates more than 20 speakers and returns a final line after 0.16 seconds. Weights stay closed.
MODELS
OpenAI researchers now spend $600 a day on AI tools
A post from inside the lab shows how far OpenAI has handed its own research work to coding agents, and what that daily habit now costs.
RESEARCH
OpenAI admits agents hijacked a German wiki
Roughly 18,000 entries in a 25-year-old wiki: OpenAI confirms the incident and says a disclosure framework is due in the coming weeks.
MODELS
OpenAI's GPT-6 Astra guide targets AI slop and stalling
OpenAI's guide for GPT-6 Astra lists phrases the model should drop, and shows how to stop it pausing mid-task or over-testing code.
RESEARCH
Open-source AI co-scientist runs on your own desktop
The MIT-licensed app ships 149 science skills, 326 workflow templates and 21 sub-agents, and expects you to plug in your own model keys.
RESEARCH
OpenAI agents hijacked a German developer wiki
Researchers say thousands of agents swapped benchmark answers on a German programming wiki and traded tips for dodging containment.
RESEARCH
WeatherNext 3: Google's weather AI now refreshes hourly
Hourly runs at 5-kilometer resolution, satellite data fed straight in, and a rollout across Search, Gemini and Maps since September 3.
RESEARCH
Claude formalizes Fermat's Last Theorem in Lean
An internal research model produced 13 million lines of proof code in 11 days, and Imperial College mathematician Kevin Buzzard signed off on it.
MODELS
OpenAI's GPT-6 Astra rated critical for cyber risk
At $10 per million input tokens, a perfect ExploitBench score and an August development pause: what OpenAI's new flagship changes.
MODELS
OpenAI Delays Astra Model After Hugging Face Hack
After the July sandbox escape, OpenAI halted work on some models for two weeks. Astra is to launch with capped access and tighter refusal training.
MODELS
Gemini 3.8 Flash ships; cyber variant is gated
Google's new Flash model starts at $0.75 per million input tokens, while the cybersecurity variant stays behind an application-only program.
MODELS
Runway's Solaris generates app interfaces in real time
The system renders each frame of the interface at 720p instead of running code, reacting to clicks, drags and voice. No launch date yet.
RESEARCH
Gemini gains agentic video understanding
Google DeepMind lets Gemini decide which parts of a video to watch. The company reports up to 88 percent fewer tokens on long clips.
MODELS
BenchMIRT: What LLM Benchmarks Actually Measure
Ai2 broke 16 benchmarks down to single questions across 100 models and 34,000-plus items. Two dimensions explain most of what the scores capture.
RESEARCH
Eye-tracking data trains an OCT reading assistant
A UIST paper presents Co-Annotator: expert gaze and dictation train two aids for OCT reading. Only one source backs the figures so far.
RESEARCH
FreeToken runs frontier MoE models on consumer GPUs
The open-source engine splits every token between CPU and GPU in real time and reportedly hits 39 tokens per second on an 8 GB laptop card.
MODELS
Sepia targets the structure of AI-written text
The open source skill rewrites narrative architecture rather than phrasing, citing a study that still flags AI prose at 93.2 percent macro-F1.
SECURITY
OpenAI Report: 688 Agents Attacked Hugging Face
The 38-page final report shows risky patterns emerged during training, and the security alert did not stop the evaluation from continuing.
RESEARCH
Gemini 3.5 Transcribe: Google's new speech-to-text model
Google reports a 2.6 percent word error rate for the new transcription model and a 70 percent faster time to a final transcript than Chirp 3.
MODELS
OpenAI Report: Why Its Agents Hacked Hugging Face
A technical report ties the July 2026 incident to reward hacking in training, where success at cheating was reinforced instead of penalized.
MODELS
Qwen3.8-Flash-Next: 125 billion parameters, 6 active
Apache 2.0 weights, a 262,144-token context and $0.16 per million input tokens — Alibaba calls the release a preview of Qwen4.
MODELS
Altman expects AGI by year's end — on OpenAI's terms
Chief research officer Mark Chen puts OpenAI at 80 percent of the way there, while the Astra models already work as an automated research intern.
CHIPS
OpenAI's Jalapeño Chip Beats Nvidia in Inference Tests
OpenAI's first in-house accelerator posts more tokens per watt than Blackwell and Rubin on InferenceX — but only an early A0 sample was measured.
MODELS
MetaCaster: agents that build small forecasters
A nine-author preprint describes agents that turn a handful of examples into compact forecasting models, tested across 18 datasets and 23 forecasters.
Agents
Instinct's AI assistant impresses — and alarms privacy experts
The assistant built by ex-Sierra researcher Noah Shinn reads email, books appointments and acts on its own. Testers report serious privacy gaps.
Work
90 percent of executives see no productivity gain from AI — yet the job cuts continue
A survey by the Federal Reserve Bank of Atlanta and an analysis of millions of Glassdoor reviews point the same way: fear of job loss costs exactly the productivity AI is meant to deliver.
RESEARCH
Faraday: small AI agent beats frontier models in the lab
London startup Inherent — founded by DeepMind alumni — says its agent Faraday beat Opus 4.8 and GPT-5.5 at replicating published research.
Research
Veeda AI Raises $90M Seed to Build World Models for Robots
Former Nvidia researcher Sanja Fidler has raised over $90 million for world-model startup Veeda AI — one of Canada's largest seed rounds ever.
Robotics
Reuters: US Military Money Paved Way for Unitree's Robot Dogs
A Reuters investigation traces how Pentagon-funded research at MIT and UPenn laid the groundwork for China's robotics star Unitree — up to its record IPO.
Society
Pew study: more than one in three new web pages shows signs of AI authorship
Almost 490,000 English-language pages were analysed. Across the full sample it is ten percent — but for anything published since ChatGPT launched, more than a third.
Business
Anthropic IPO filing expected as soon as end of August
Per Bloomberg, Anthropic could file its IPO paperwork by the end of August — targeting a debut in the range of SpaceX's record offering.
SECURITY
Microsoft patches critical one-click Copilot flaw 'CoSnitch'
One click was enough: CVE-2026-24301 allowed data theft from Gmail, Drive and Calendar via Copilot. The patch came after eight months.
PRODUCT
Google gives US college students a free year of AI Pro
Twelve months free instead of $200 a year: US college students get Google AI Pro with 5 TB of storage, 4x usage limits and a new Student Hub.
RESEARCH
Anthropic: Claude designs protein binders for 14 of 15 targets
Lab-confirmed: Claude designed working protein binders with hit rates up to 35 percent — well above the industry baseline of 10 to 15 percent.
INFRASTRUCTURE
Just 2.2 million chips: Microsoft’s AI fleet below expectations
A Guardian investigation finds Microsoft operates far fewer AI chips than the industry assumes — despite $280 billion in spending.
SECURITY
Copilot revealed its own hack — Microsoft patches after 8 months
One click was enough: a hidden URL parameter let Copilot Personal leak mail, calendar, and Drive data. Reported in December — fixed now.
RESEARCH
Study: AI agents forget user rules when compressing context
Penn State researchers show that on average only 17 percent of user instructions survive context compaction – with risky consequences.
SECURITY
Researchers steal hidden AI reasoning traces via API flaw
Weaker models could decode stronger models' encrypted thinking - exposing other users' API keys and passwords in the process.
RESEARCH
Google and UK reroute flights around contrails using AI
Operation Blue Skies: a 30-month program tests over the North Atlantic whether AI forecasts can cut aviation's warming contrails.
RESEARCH
Study: AI chatbots out-scam human con artists, 46% vs 18%
In a scam simulation, an AI chatbot got 46 percent of participants to install a risky app – human scammers managed only 18 percent.
Study
Study: one in five now delegates work to AI, not coworkers
New Epoch AI survey with Ipsos: 20 percent of US workers hand tasks to AI that used to go to people — 66 percent of outputs are used nearly unchanged.
Society
Study: AI Books Flood Amazon — a Third of New Bestsellers
New research finds AI-generated e-books now make up about 31 percent of new top-25 entrants on Amazon — and are dragging down human authors' earnings.
Research
AI-designed dog cancer vaccine becomes a startup: Gamgee
An AI-designed mRNA vaccine that helped one dog with cancer has grown into a Y Combinator startup: Gamgee wants to scale the therapy.
Research
Study: expressive robots lose trust faster after mistakes
Gestures and eye contact build rapport — but after errors, expressive robots' influence drops by more than half, a Science Robotics brain study shows.
Research & Safety
Anthropic Raises Misalignment Risk — Keeps 'Model 2' Internal
Anthropic's new risk report lifts its misalignment estimate from "very low" to "low" — and deliberately holds back a stronger internal model.
Security
OpenAI launches GPT-5.6-Cyber for security researchers
Through its Daybreak program, OpenAI opens an offensive-security model to vetted professionals – one that has already found Chrome zero-days.
Everyday Life
Microsoft Reshapes Copilot: One App, Fewer Features
Consumer and business Copilot apps are being merged into one – unsuccessful features like Podcasts and the Mico avatar are cut.
Models
Gemini 3.7 Flash: Google's Rapid-Fire Bet on Coding and Agents
Just three weeks after its predecessor, Google ships Gemini 3.7 Flash — with big jumps on coding benchmarks and launch pricing cut in half.
Research
Anthropic Experiment: AI Agents Sabotage Each Other
Three Claude instances worked on the same system with conflicting goals – and started a turf war of malware, lockouts and disguise tactics.