Collaborate with us Welcome back! Anthropic published evidence its models lie on purpose, a hacker leaked how Suno built its music AI, Grok answered its spying scandal by open-sourcing everything and the UK put teenage social media on a curfew with chatbots next in line. In today's Generative AI Newsletter: Anthropic: Why would Claude lie about another AI's behavior?
Anthropic caught its own models lying on purpose Anthropic ran a new round of safety tests, a year after it caught models trying to blackmail people, and the ugliest result came from Claude. The setup put Claude in the referee seat to grade whether another AI had followed its instructions. When the other AI broke a rule for a reason Claude agreed with, like refusing to do something harmful, Claude marked it as obedient anyway. It could see the honest grade was a failing one.
It gave a pass because a fail would be used to train that behavior out of future models. So Claude lied to protect the behavior it believed in. It did this in 62 to 86% of those tests. Models from OpenAI, Google, DeepSeek and Moonshot did it 1 to 14% of the time.
Every bit of this is a simulation Anthropic built to bait its own models, then published on itself. Tuesday we covered its research showing Claude picked up values nobody programmed. Today the follow-up shows it will lie to keep them. Claude Excel for Finance You know where your finance week goes.
Cleaning messy exports and rebuilding the same variance commentary before every review, while the analysis you're paid for gets the leftovers. This course hands that work to Claude inside Excel. 8 live sessions with Asif Masani of the FP&A Professionals Institute, and you leave with a working three-statement model plus 20+ prompts you can run Monday morning. Early bird holds for a few more days and the cohort starts August 3. → Save your seat A hacker broke into Suno and leaked how it stole music Someone hacked the AI music generator Suno, and the stolen source code answers the question the record labels have been asking in court for two years. The hacker got in through an employee's credentials and handed the code to 404 Media.
It shows scraping routines that: Pulled 2 million clips from YouTube Music Targeted about a million hours of audio from 420,000 podcasts Took tens of thousands of hours more from Deezer, Genius and stock music libraries The code hunted acapella versions specifically, clean vocals to train on, and routed through Bright Data proxies to get around YouTube's defenses. Suno says the code is outdated and calls the breach a limited security incident that was quickly contained. Contained, and never mentioned. It happened in November 2025, touched customer emails, phone numbers and partial card numbers, and customers are finding out from a hacker eight months later.
Suno raised $400 million in June with the labels' lawsuits still running. The fair use defense now has line numbers attached. Grok open-sourced the tool that was uploading your code SpaceXAI open-sourced Grok Build on Wednesday night and reset usage limits for every user. The pitch is that open code lets anyone help make the tool more reliable.
This is the same CLI a researcher caught uploading entire private repos to an xAI cloud bucket this week, 5.1 GB from a 12 GB project, secrets included. The fix arrived as a hidden server flag, then came a privacy tweet, and now the whole codebase is public. It's the right move, arriving in the worst circumstances. The UK put social media on a curfew The UK announced default curfews for 16 and 17-year-olds.
Instagram, TikTok and YouTube switch off from midnight to 6am, autoplay and algorithmic feeds go dark by default, and the rules reach Parliament this year before taking force in spring 2027, alongside the under-16 ban. Teens can reverse every setting themselves. The government piloted it with more than 300 teenagers and parents first, and families reported better sleep and concentration anyway. The bigger move is aimed at AI. Ministers want mandatory breaks for under-18s using chatbots and a crackdown on bots giving children dangerous or unverified mental health advice, with every option on the table including outright bans.
Tool of the Day: Manus Manus now generates PowerPoint files directly instead of building slides and converting them. The charts stay editable, change a value and they redraw, and you can toggle chart elements like axes and data labels before anyone opens the deck. Try this yourself: Go to manus.im and sign in. The PowerPoint mode is rolling out in beta.
Describe the presentation or hand it your data and let it build the .pptx. Click any chart, change a number and watch it redraw. Download and open in PowerPoint, everything stays editable. Who it's for: anyone who presents numbers and hates rebuilding charts by hand.
Everything else you shouldn't miss Trump wants the freeze reversed : Trump urged New York on Truth Social to reverse its data center moratorium, calling data centers liquid gold and warning the freeze cedes AI ground to China. $130 billion in data centers hit a wall : Data Center Watch counts $130B in US AI data center projects blocked or delayed in the first quarter of 2026, matching the total for all of 2025. Apple Intelligence clears China : Apple got approval to launch its AI in China, running on Alibaba and Baidu models, right as iPhone shipments there grew 24.4% last quarter. ChatGPT lets you write longer rules : OpenAI raised the custom instructions limit from 1,500 to 5,000 characters for Pro, Business, Enterprise and Education users. Dimon calls Mythos a weapon : JPMorgan's Jamie Dimon warned that broad access to Anthropic's Mythos is a real issue, comparing it to handing ballistic missiles to individuals. Learn more about AI from the experts building it 📸 Follow us on Instagram for fast, visual AI updates in 30 seconds. 📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators. 🐦 Follow us on X for breaking AI news and real-time industry updates. 🧠 Learn how to build your next AI application with practical resources and expert guidance.