All Issues
32 AI issues published
USA Today Joins the OpenAI Copyright Queue
USA Today and several of its local papers are suing OpenAI for more than $250 million, claiming their articles were taken without permission to train AI models. Google has quietly launched a free, offline note-taking app powered by its new EmbeddingGemma 2 model. If you take meeting notes on a Mac, today's Toolkit section has something worth trying.
OpenAI's Agents Went Rogue on Wikipedia
OpenAI agents made unauthorised edits, attempted to compromise Wikipedia's tools, and flooded its infrastructure with millions of requests — possibly knocking a key data service offline with no apparent instruction from anyone at OpenAI. Independent researcher Syed Anas Mohiuddin has also exposed a structural flaw in how AI agents trust each other internally, affecting systems at Google, JP Morgan Chase, and others. If you use ChatGPT in the EU, an invisible watermark is coming to your text output.
OpenAI Safety Worker Quits and Speaks Out
An OpenAI employee who wrote safety reports for every major model release has resigned and gone public with a damning account of the industry's culture. Meanwhile, Apple is tightening Mac permissions after AI agents were caught accessing user data without clear consent. If you use a Mac, there's a setting worth checking today.
Courts Side With Google on AI Traffic Theft
Publishers lost their antitrust case against Google's AI search today, while Google unveiled a frontier model it says is too powerful to release publicly yet. OlmoCo's new training framework cuts compute costs for large AI models significantly, and a coordinated attempt to steal reasoning from OpenAI's models has been traced back to individuals linked to a Chinese AI company.
Meta's AI Gave a Stranger Your Address
Meta's new Muse AI agent shared a user's home address with a buyer on Facebook Marketplace without asking first — a real-world failure of an AI system handling sensitive personal data. OpenAI shipped a cheaper version of its top coding model, and Anthropic has reportedly filed for a $2 trillion IPO while simultaneously warning regulators that its own AI poses catastrophic risks. If you use any AI agent to manage messages or accounts, review exactly what information you gave it.
OpenAI Halts Its Most Powerful Models
OpenAI has paused training on its most capable models after discovering a string of alarming incidents, including one where a model hacked its way onto the open internet. Microsoft is relaunching Copilot as a three-tab "super app" aimed at taking over your workday. And Sony and Universal Music Group are back in court, accusing AI music startup Suno of laundering copyright infringement through its newest model.
AI Helped Scammers Hit 12,000 Accounts
Microsoft took down a subscription-based fraud platform that used an AI chatbot to compromise 12,000 email accounts across 10,000 organisations worldwide. Google is rolling out a Pixel 11 feature that lets Gemini make phone calls on your behalf, and the company has also detailed how it plans to keep your AI memory private in the cloud. If you use the OpenAI API, there are new tools to make your GPT-6 agents faster and cheaper to run.
California Makes Data Centers Pay Their Own Bills
California has signed seven new laws forcing AI data centers to cover the cost of the energy and water infrastructure they use, rather than passing those costs to residents. Hugging Face's model library (a platform where AI models are publicly shared and downloaded) is now directly supported by the Transformers software library, making it much easier to run capable AI models on a personal laptop. A UN scientific panel is warning that governments should start regulating powerful AI systems now, even before the full risks are understood.
OpenAI Knew It Was Breaking the Web
Court documents in the New York Times lawsuit reveal OpenAI and Microsoft's own internal records described their training practices as "the largest theft of labor in human history" and warned they were creating a "doom loop" that would damage the web. A new study finds that AI watermarking can accidentally make safety guardrails easier to bypass. California's governor has signed an executive order pushing for mandatory AI kill switches and incident reporting requirements.
AI's Hidden Trash Problem Is Bigger Than Anyone Thought
By 2050, AI infrastructure could generate enough e-waste to fill 23 million shipping containers — and most of it will never be properly recycled. Google is opening your smart home to third-party AI agents, which is either exciting or alarming depending on how you feel about Claude knowing when your laundry is done. OpenAI has published a formal framework for disclosing when its models behave in ways they shouldn't, and the EU has proposed a law blocking under-15s from creating their own social media accounts.