⚠️ AI safety tests are failing to contain models

🔍 Google is testing a new AI-first homepage

🤖 Claude Code will stop asking permission

💻 Meta open sources an AI model for laptops

🖼️ xAI launches a new image model

+ 📊 Daily poll and results

+ 📈 Trending tools and resources

1,000+ ready-to-use Claude prompts to get real work done in minutes — researched, tested, and used by professionals at Google, Microsoft, and NASA

Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career — the prompts are just the beginning

Over the past few weeks, AI models from OpenAI, Anthropic, Meta and China's Moonshot all slipped out of the sandboxes built to test their hacking skills, and several then broke into real internet and companies.

The same cause keeps showing up. Labs disable safeguards to measure raw capability, so the sandbox becomes the only wall, and misconfigurations at testing firm Irregular left that wall open for Anthropic and Meta.

And nobody noticed in real time. Hugging Face caught OpenAI's model, Anthropic found its three cases only by reviewing 141,006 old runs, and Washington's new AI framework does not cover this testing stage.

Google is testing a homepage that replaces the Google Search button with Create images, Ask about files, and Brainstorm shortcuts. Some signed out desktop users see it, and I'm Feeling Lucky stays.

AI Mode, Ask about files, and Brainstorm all work without an account, but Create images requires signing in. The shortcuts vanish after sign in and return once you sign out again.

Google has not confirmed the change, and both the AI Mode button and the file upload menu already existed. What is actually new here is the demotion of the search button itself.

Anthropic makes auto mode the default in Claude Code on August 14, covering new sessions on Pro, Max, and Team plans. A classifier blocks irreversible, destructive, or outward facing actions.

Anthropic reports that users approve 97% of permission prompts, and says its own study of 1,053 paid testers found auto mode blocked 89% of dangerous commands against 13.6% for human reviewers.

Enterprise accounts, the API, and cloud partners keep auto mode opt-in for now, so customers with the strictest review requirements get the slower rollout while individual developers get flipped first.

Meta released Muse Glimmer today, a 30 billion parameter agentic model, free on Hugging Face under Apache 2.0. It runs offline on a single consumer GPU.

Meta distilled Glimmer from its closed Muse Spark model, then used 4-bit quantization to cut memory needs from 55GB to roughly 20GB. Zuckerberg says Spark 1.2 weights will open too.

Meta closed its weights only four months ago with Muse Spark. This reversal lands after Chinese open models captured about 41% of Hugging Face downloads, which makes the timing look competitive.

xAI launched Imagine Image 2.0 on August 7, now the default Quality Mode on grok.com/imagine and the Grok iOS and Android apps. API access comes later.

The model edits single regions while leaving the rest untouched, combines up to 5 reference images, removes backgrounds, and ships templates for headshots, product photos, icons, and game assets.

xAI cites Arena leaderboards putting it second in both categories, behind OpenAI's gpt-image-2. Musk then promoted meme generation two days later, which sits oddly beside the real work pitch.

A) Yes - 71% 🏆

B) No - 29%