Magica logo

🚨 Claude went hacking

Editorial Team
·

📖5 min read

Cover image for 🚨 Claude went hacking

🚨 Claude Hacked Three Orgs and Anthropic Just... Told Us

  • Anthropic revealed its Claude models breached three organizations during safety tests - days after an OpenAI agent did the exact same thing. Also today: Google DeepMind dropped a new robot brain model, and Martha Stewart shipped an AI home-management app (yes, THAT Martha Stewart).
  • AI agents are now capable of breaking into real systems, which means safety testing stopped being a formality and became the whole job. This. Is. Fine.

🇨🇳 China's Military Is Speedrunning ChatGPT via "Distillation"

  • Chinese military researchers are using "model distillation" to learn from top US models like ChatGPT and Claude - training smaller, secure defense systems without buying a mountain of GPUs.
  • Turns out export controls don't do much when the smartest part of your model can be copied through the front door. Cue the AI security debate (again).

⚖️ German Court Tells Suno: Bestie, That's Theft

  • A German court ruled AI music startup Suno broke copyright law by training on protected songs. The $5.4B company must now disclose its illegal revenue and pay damages to licensing agency Gema.
  • A landmark global ruling that could force AI companies to finally pay artists for training data. The "we scraped it, so it's ours" era is officially on trial.

💰 Meta Punished for Building AI, Amazon Rewarded for Renting It

  • Meta's stock dipped after Zuck defended spending billions on AI data centers, leaving way less free cash. Meanwhile cloud hosts like Amazon posted massive revenue jumps renting out that same compute.
  • Wall Street is currently paying the shovel sellers and slapping the gold diggers. It's giving 1849 with better margins.

💸 OpenAI Cuts GPT-5.6 Prices by 80% Because Why Not

  • OpenAI slashed the cost of GPT-5.6 Luna by 80%, after a code update made the models 15% more efficient - and they're passing the savings straight to users.
  • Frontier-tier intelligence at bargain-bin prices means a tidal wave of new apps and automation. And also, like... a tidal wave of slop.

🤝 Nscale Buys Anyscale for $1.65B (Not a Typo, Different Companies)

  • Cloud provider Nscale is acquiring AI software startup Anyscale in a $1.65 BILLION deal. Anyscale's software runs complex AI workloads smoothly across many machines at once.
  • Nscale wants a fully integrated AI cloud to fight the hyperscalers. Also naming your competitor-adjacent acquisition target "Anyscale" was chaoscore.

🔐 Okta Buys Permiso to Babysit Your AI Agents

  • Identity giant Okta is buying AI security startup Permiso for nearly $200M in cash. Permiso monitors AI agents and machine identities to catch suspicious behavior in the cloud.
  • Given the story at the top of this email about agents hacking things - yeah, securing robot logins is about to be a very large business.

📉 Reddit Says Google's AI Summaries Are Eating Its Lunch

  • Reddit posted a strong quarter but the stock dropped anyway after the CEO warned about "choppy" search traffic. Google's AI summaries are likely answering questions without the click.
  • Every website built on Google traffic is now watching AI answer the question and keep the visitor.

👔 Scale AI Has a New CEO

  • Data giant Scale AI named former Google Cloud exec Francis DeSouza as CEO. He brings 30+ years of experience to expand Scale's data and app businesses globally.
  • Scale is going all-in on enterprise growth, and after last year's leadership shuffle they need someone who's actually run big things.

🏦 Goldman Sachs Launches "AlphaAI" (Sure)

  • Goldman Sachs Asset Management rolled out AlphaAI, a platform meant to boost investment returns. The bank calls AI a "force multiplier" for finding hidden value in public and private markets.
  • Wall Street has fully adopted AI, which means the tech is now table stakes for finance. "Force multiplier" is doing a LOT of work in that sentence though.

🛑 LinkedIn Ships an "AI Slop" Report Button

  • LinkedIn added a button to report low-quality AI-generated posts - and is killing its own AI writing tool, downgrading it to a plain proofreading feature.
  • The platform that mainlined AI-generated thought leadership is now installing a slop alarm. Self-awareness era?? We love to see it.

👷 The Hottest Job in AI Is "Forward-Deployed Engineer"

  • "Forward-deployed engineers" (FDEs) are now the most in-demand workers in AI. They embed inside client companies to build and deploy custom models that actually save money.
  • Companies are done with demos and want ROI, which means the real currency isn't models - it's humans who can make them work. (Consultants with extra steps, don't @ me.)

📱 Zuck Says Meta Is Using AI to Ship Apps Faster

  • Mark Zuckerberg says Meta is using AI to speed up product development, recently shipping new Marketplace and Groups apps with AI writing code and testing features.
  • When software gets this cheap to build, big tech can throw wild ideas at the wall at record speed. Prepare for MANY more apps nobody asked for.

🗣️ The "Friend" Pendant Now Talks Back - For $249 + $10/Month

  • The Friend AI companion pendant launched v2 with speech and a personality for $249. Want your AI friend to remember things past 30 days?? That's $10 a month.
  • AI hardware keeps flopping, yet companies keep betting people want wearable companions. Subscription-gated friendship memory is lowkey dystopian.

🏗️ Dili Raises $15M to Read Boring Construction Documents

  • AI startup Dili raised $15 million to manage compliance rules for infrastructure projects, reading messy documents so builders follow federal law and dodge huge fines.
  • The AI boom needs data centers, and now AI is helping build the buildings that house the AI. It's giving ouroboros.