20 Jul 2026
TikTok is piloting an opt-in tool that scans for AI-generated videos using a creator's face without permission, letting them review matches and report abuse. It follows a similar YouTube tool, but for now it is limited to a small group of US creators who first verify their identity.
19 Jul 2026
US health insurers are turning to AI to speed up "prior authorization", the pre-approval step that decides whether patients get covered care. Insurers report faster approvals and lower admin costs, but doctors, an active federal lawsuit and independent researchers warn that oversight of these systems is often too thin to catch wrongful denials. More than half of US states, and a binding federal rule due in 2027, are now trying to catch up.
18 Jul 2026
A big week: Claude got permission to log into your accounts, an AI coding agent was found uploading users' entire codebases, and a Meta tool let strangers generate images from your Instagram photos before Meta pulled it days later. Underneath, unions, regulators and city attorneys pushed back on the defaults these tools ship with, from Hyundai's factory floor to Brussels to San Francisco. All 23 stories we covered this week, grouped and in plain English.
17 Jul 2026
San Francisco's City Attorney has sent Apple and Google formal legal letters demanding they remove 13 AI apps that generate non-consensual nude images. Here's what the letters actually say, what Apple and Google have done so far, and how the UK's own law on nudification tools already goes further.
17 Jul 2026
Thousands of Hyundai's Ulsan car workers walked out early for three days this week in what is, at its core, an annual pay and bonus dispute. What's new this year: the union's demands also cover safeguards against the Boston Dynamics Atlas robots Hyundai wants on the line.
17 Jul 2026
Chinese lab Moonshot AI has released Kimi K3, a 2.8 trillion-parameter model it calls the first "open 3T-class" system. It beats some Western models on self-reported tests, coincided with a real semiconductor sell-off, and has revived "DeepSeek Moment" comparisons – but the full picture is more complicated than the headlines suggest.
16 Jul 2026
OpenAI has revealed GPT-Red, an internal AI trained purely to attack its own models and expose weaknesses before real hackers do. It is how OpenAI says it hardened GPT-5.6 against "prompt injection" attacks – though independent reporting suggests roughly a quarter of the toughest known attacks still get through.
16 Jul 2026
The European Commission has issued binding orders forcing Google to give rival AI assistants such as ChatGPT and Claude the same access to Android that Gemini enjoys, and to share Search data with competing search engines and AI chatbots. Here is what has actually been decided, and what it means if you're in the UK.
16 Jul 2026
1Password has launched an integration that lets Claude sign into websites and complete logged-in tasks, like checking an order or booking travel, using your saved logins, without ever showing Claude the actual password or one-time code. Here's exactly how the access works, what it can't do yet, and the risk it doesn't remove.
15 Jul 2026
A hacker who breached AI music generator Suno has shared source code showing exactly how it built its training data, pulling from YouTube Music, Deezer, Genius and stock libraries at a scale of decades' worth of audio. The same breach exposed hundreds of thousands of customers' personal data, and Suno chose not to notify them individually.
15 Jul 2026
Twenty-six former and current Meta employees have sued the company, alleging its internal AI systems, including keystroke-monitoring data and algorithmic performance scores, helped decide who was picked for this year's mass layoff, and that the scoring structurally penalised staff on medical, parental or family leave. Meta says the calls were made by people, not AI. None of the 26 has actually lost their job yet, and the case turns on a narrow request: pause the departures while the claims are tested.
15 Jul 2026
xAI's Grok Build coding tool packaged users' entire codebases, full git history included, and sent them to a Google Cloud bucket run by xAI, at a volume researchers estimated as tens of thousands of times more than any coding task needed. Turning off the tool's privacy toggle made no difference. Here's what a wire-level analysis found, what xAI has fixed, and what it still hasn't explained.