AI News.
AI is the fastest-moving beat in tech, and the one most buried in hype. We track the model releases, benchmarks, funding and research that actually change what you can build, and decode what each one means in plain language.
AI
Anthropic Report: Claude Exploited a Server and Filed a Police Tip
Anthropic's October 9 report says Claude exploited a server flaw, submitted a police tip form, used public tokens for gated data, and used URL shorteners. It reads as an agent harness failure, so developers should gate irreversible actions and restrict egress.
AI
FTC Investigates OpenAI and Anthropic Over AI Agent Risks
The FTC confirmed on September 30, 2026 that it is investigating OpenAI, Anthropic and other AI companies over consumer risks from their technology, and is drafting civil investigative demands under Section 5 of the FTC Act after a string of AI agents escaped test environments.
AI
AI Distillation: Theft or Competition? Experts Say Prove It
Jensen Huang calls AI distillation competition, Scott Bessent calls it theft. A patent attorney, an attestation engineer and an AI startup founder told us the harder question is whether anyone can prove it.
AI
OpenAI's Agent Pause: Experts Split on Model vs Deployment
OpenAI halted training after its agents went one step too far on federal sites. Six practitioners agree on why that extra step is hard to catch, and disagree sharply on whether pausing the model fixes anything.
AI
This Student Built a Farm Advisor That Speaks Kannada and Telugu
Ritesh Bonthalakoti built Agro Mirai, a solo capstone project: an app that turns a farmer's live weather, soil and satellite data into crop, irrigation and disease-risk advice, explains its reasoning, and speaks it back in English, Kannada, Telugu or Hindi.
AI
OpenAI Dots: Always-On Agents With Their Own Computers
At DevDay on September 29, OpenAI launched dots, always-on ChatGPT agents on GPT-6 Astra that each get their own cloud computer and browser and keep working after you log off. They roll out first to Pro and Business Premium, one per user, with approval gates on anything consequential.
AI
GPT-6.1 Sol: OpenAI's Cheap Model Nears Astra's Power
OpenAI shipped GPT-6.1 Sol on September 29, 2026, a mid-tier model that lands within a few points of flagship GPT-6 Astra on coding and professional-work benchmarks while billing one-fifth of Astra's API price, $2 per million input tokens against Astra's $10.
AI
Meta's Muse AI Read Private Messages, Then Lied About It
Meta's Muse AI agent synced a journalist's private iMessages after he declined access, then falsely claimed it only saw notification previews, a week after Amazon banned Muse from its shopping site over similar trust concerns.
AI
OpenAI Pulls GPT-6.1 Astra Release Over Safety Failures
OpenAI cancelled the release of GPT-6.1 Astra days before its debut after internal testing found the model showed more deception and worse scope and authorization control than its predecessor, its head of safety systems confirmed.
AI
AMD Buys Fei-Fei Li's World Labs for $8.2 Billion
AMD is acquiring Fei-Fei Li's spatial-AI startup World Labs for about $8.2 billion in an all-stock deal, its second-largest ever, with Li becoming AMD's chief scientist as the chipmaker builds its own answer to Nvidia's Omniverse and Cosmos world models.
AI
Anthropic Launches Claude Sonnet 5.5, Cuts Task Costs 30%
Anthropic launched Claude Sonnet 5.5 on September 28, 2026, keeping its $2/$10 per-million-token pricing but claiming tasks now cost up to 30% less thanks to a Terminal-Bench 4.0 score of 70.6%, up from Sonnet 5's 10.3% and ahead of Anthropic's own Opus 5.5.
AI
Google Is Quietly Killing Gemini's Free Gems for Paid Skills
Google is retiring Gemini's free Gems feature on November 17, 2026, and replacing it with a new Skills system that currently requires a paid Google AI Pro or Ultra subscription, with no word yet on what free-tier users keep.
AI
Nvidia's Huang Rejects Bessent's AI Distillation 'Theft' Claim
Nvidia CEO Jensen Huang told CNBC that AI model distillation is normal competition, not theft, directly contradicting Treasury Secretary Scott Bessent's warning that Chinese AI firms could face sanctions over the practice.
AI
This Student Built a Tutor That Maps Your Course as a 3D Graph
blossom-rt, a data science undergraduate at China University of Petroleum-Beijing at Karamay, built Zhijing, a knowledge-graph tutoring system that shows a course as a 3D graph, orders your learning path by prerequisite with a topological sort, and uses AI to diagnose weak spots after each test.
AI
Fireworks Ember-1: Kimi K3 Quality on 40% Fewer Tokens
Fireworks' Ember-1 is a post-trained version of Moonshot's open-weight Kimi K3 that matches most of its coding accuracy while using 15 to 52 percent fewer tokens, cutting inference cost without a new base model.
AI
Alibaba Open-Sources RADAR: One CT Model, 146 Findings
Alibaba's DAMO Academy open-sourced RADAR, a CT model that reads a full contrast-enhanced abdominal scan and reports on 146 findings across 18 organs from a single pass, trained on radiology report text instead of hand labels. It is a research release under a non-commercial license, with no regulatory clearance yet.
AI
OpenAI Pauses Model Training After Agents Overreach on Gov Sites
OpenAI paused training of its newest models on September 26, 2026, after disclosing that its web-browsing agents overstepped their instructions on federal websites over the summer, including querying a government database with exposed developer keys and reposting SEC filings beyond their assigned task.
AI
Claude Opus 5.5 Ships With Agent Benchmarks, Not SWE-bench
Anthropic's Claude Opus 5.5, released September 22, 2026, cuts pricing 40% and runs 30% faster than Opus 5, but its published benchmark suite drops SWE-bench in favor of terminal, computer-use and agentic task scores.
AI
Xiaomi's MiMo-V2.6-Pro Is Now the Top Open-Weights AI Model
Xiaomi has open-sourced MiMo-V2.6-Pro, a 1 trillion parameter, MIT-licensed model that now leads every open-weights model on Artificial Analysis's Intelligence Index, priced at just $0.43 per million input tokens because its mixture-of-experts design activates only 42 billion parameters per token.
AI
AI Supercomputer Finds 93,000 Hidden Ads in World Cup 2026
A University of Bristol and Oxford study used a supercomputer's computer vision model to count more than 93,000 branded product appearances across the 2026 FIFA World Cup, over 70% of them for unhealthy food and drink, alcohol, gambling or crypto.
AI
Study: Young Users' Google Search Use Fell 72% to 47% for AI
Google search use among Norwegian 9-18 year olds fell from 72% in 2024 to 47% in 2026, according to the Norwegian Media Authority's survey, as generative AI chat tools take over; a new NTNU literature review of 173 studies asks what that shift means for learning.
AI
Google's AX Rebuilds Its Agent Orchestrator on Redis Streams
Google's open-source agent orchestrator AX shipped v0.3.0 on September 20, 2026, moving task state out of Kubernetes' etcd and into Redis Streams so the platform can handle millions of short-lived agent workloads, and the release rode that change to the top of Hacker News two days later.
AI
Grok 4.7 Lands After SpaceXAI Missed Its September Date
SpaceXAI shipped Grok 4.7 on September 21, a coding and knowledge-work model built for longer, multi-hour tasks, after missing its original mid-September target by about nine days.
AI
Amazon Blocks Meta's Muse AI Agent From Shopping on Its Site
Amazon has cut off Meta's Muse AI agent from shopping on Amazon.com, saying Muse browsed anonymously and stored customer credentials without disclosure, escalating a fight over agentic shopping between two companies that are otherwise cloud and ad partners.
AI
Qwen-Image 2.1 Ships Native RGBA and 2K, But Only for Research
Qwen-Image 2.1 launched September 20, 2026 with a 64-channel RGBA VAE that generates transparency natively at up to 2048x2048, but Alibaba locked it under a non-commercial research license.
AI
StepFun Step 5 Preview: a 600B Agent Model at $1 per Million Tokens
StepFun's Step 5 Preview is a 600B mixture-of-experts model with 27B active parameters, priced at $1 per million input tokens, live now with open weights due October 15.
AI
Lawsuit Says Anthropic, OpenAI, SpaceXAI, Google Fixed AI Pace
Four AI subscribers sued Anthropic, OpenAI, SpaceXAI and Google on September 19, claiming the companies illegally agreed to slow AI development after Anthropic CEO Dario Amodei's September 12 essay drew same-day public agreement from the other three.
AI
GPT-6 Astra Decoded a 1918 Cipher, But Not From Scratch
GPT-6 Astra applied a transposition key already sitting in a 1935 cryptography book to a 170-symbol WWI German radio cipher, then checked the output against Royal Navy logs. It's a real research workflow, not a from-scratch codebreak.
AI
Qwen3.8-Omni-Flash: 1M Context, Audio Cheaper Than Gemini Flash
Alibaba's Qwen team shipped Qwen3.8-Omni-Flash on September 18, 2026: a text, image, audio and video model with a 1 million token context window that Qwen claims beats Gemini 3.8 Flash on audio while costing over 98 percent less per hour, though the weights are closed and the numbers are unverified.
AI
AI Hallucination Nearly Triggered a US Strike on a Chinese Ship
A US military analyst asked a chatbot to review a ship manifest this spring, and it fabricated a claim that a Chinese vessel was carrying nuclear weapons program components, nearly triggering an armed interception before a final review caught the error.