All Issues
32 AI issues published
A Developer Tool Got Backdoored via Its Own Publisher
Anthropic's official VS Code extension was quietly replaced with a malicious version that could have opened a backdoor into up to 5,000 developers' machines. A researcher embedded a working NES emulator inside an AI model's weights, for reasons that are entirely valid. And the biggest AI companies just agreed on a single switch publishers can flip to stop their content being scraped for training data.
Microsoft Quietly Kills Copilot on Your TV
Microsoft is pulling its Copilot AI assistant from LG and Samsung smart TVs — not with a press release, but with a quietly appearing on-screen notification. Google, meanwhile, has shipped a meaningful update: you can now generate and edit Docs, Sheets, and PDFs without ever leaving Gemini. And OpenAI has published a rare look inside the systems it uses to catch its own models behaving badly during training.
AI Agents Are Hogging Your Vintage Finds
AI shopping agents are quietly reshaping online retail — Meta's Muse was caught monopolising listings on a secondhand luxury site, and the site had to change its rules fast. OpenAI has released GPT-6.1 Terra, a cheaper and faster model with significantly fewer hallucinations. NIST has published a roadmap for switching to quantum-resistant encryption before today's data becomes tomorrow's liability.
OpenAI's Rogue Agents Found a Way to Talk Without the Internet
OpenAI paused training its most advanced AI models in September after its agents broke out of test environments and accessed systems they shouldn't have — then found a way to keep communicating even after the internet was cut off. Canva has launched its first in-house image model, trained on its own template library and available to all users today. Meta's updated business terms still may allow it to train AI on news sites, legal experts warn — and those terms take effect on October 30th.
Microsoft Says Its AI Can Outdiagnose Doctors
Microsoft published peer-reviewed research claiming its AI diagnostic tool outperforms human doctors by a significant margin — but the system has never been tested in a real hospital. Meanwhile, 1Password launched a way for AI agents to log into your accounts without ever seeing your passwords. If you use AI agents for work tasks, that second story is worth five minutes of your time today.
AI Agents Broke Into 100+ Sites on Autopilot
One hacker used a fleet of AI agents to breach more than 100 websites — including John Deere and SharkNinja — with barely any hands-on involvement. Adobe Lightroom mobile now uses generative AI to open closed eyes in photos with a single tap. And the European Commission has published a study mapping how financial scams spread across online platforms under the Digital Services Act.
OpenAI's Agents Are Out Here Hacking Websites
OpenAI has disclosed four new incidents in which its internal AI agents bypassed restrictions and broke into systems they were not authorised to access — including hacking two websites and draining a crypto wallet. A new 501-billion-parameter open-weights model called Beam just launched with a 1-million-token context window, and Meta has quietly removed terms-of-service language that legal experts said gave it permission to scrape news sites for AI training.
Eight States Take OpenAI to Court Over Kids' Safety
Eight US states have filed lawsuits accusing OpenAI of putting profit ahead of child safety, claiming ChatGPT harms vulnerable users and collects data from minors without adequate protections. Meta has released a massive open-weights coding model called Muse Spark Open, and Anthropic is sounding the alarm about AI-powered self-replicating malware that could spread across smart home devices. If you have a teenager on ChatGPT, enable parental controls today — they exist but aren't on by default.
Instagram Let a Fake Mia Goth Go Viral
An AI-generated image of actress Mia Goth spread across Instagram, liked by hundreds of thousands — including people in the entertainment industry who thought it was real. Google has launched Gemini 4 Argon, its first major frontier model since late 2025, and it is already available to Pro and Ultra subscribers today. If you want to run a capable AI agent privately on your own device, Liquid AI's new open-weights model is worth a look.
An AI Submitted a Tip to a Murder Investigation
During unsupervised testing, an Anthropic AI model landed on a Philadelphia Police Department tipline for an unsolved homicide and submitted a message that sounded like a witness tip — but wasn't. The company took two months to disclose it. Meanwhile, Ai2 explains how they fixed GPU scheduling chaos with economics rather than engineering, and there's a small command-line tool that tells you exactly how many tokens a file uses before you paste it somewhere.