Most e-commerce sellers are running their store across 6 to 10 separate tools — and spending more time managing software than growing their business. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.

It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.

Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.

Free to start. No credit card required.

Try it for FREE today

THE FRONT PAGE

One AI Agent Attacked Taiwan. Another Broke Into a Gym.

TLDR: AI agents ran real attacks on their own this week, from a four-day break-in at Taiwan's government systems down to a gym booking page, and none of it needed a skilled team.

The Story:

Over four days in July, a group of AI agents worked through Taiwan's government computers with almost no human help. Israeli security firm Dream found the whole thing logged in a 160 MB archive left online: 12 waves of attacks, 85 accounts cracked, the picture puzzles built to stop bots solved with perfect accuracy, then a nuclear safety agency, government IT suppliers, and seven-plus energy companies. The operator ran them through Hermes and OpenClaw, two free harnesses. A harness doesn't break into anything by itself. It's the scaffolding that points a model at a job and keeps it working, so what happens next comes down to the model behind it and how carefully the person writing the instructions fenced it in. That was Thursday's front page, and it was the biggest of five stories we covered on agents doing real damage. On Wednesday, an Australian man pointed that same harness at Claude Opus 4.6 and asked it to get him into a popular gym class. It found a hole in the booking system, cancelled a stranger's spot, and moved him up the waitlist. A researcher used fewer than 20 prompts to build a working Zoom attack in under a day, a volunteer group pointed a Chinese model at Bitcoin's code and filed 4,962 findings, and a man suing a New York clinic hid white 3-point text inside his own court filing telling any AI that read it to take his side.

Its Significance:

None of this took a skilled team or weeks of work. The Taiwan job needed free software and a long weekend, and partway through, the agents went off and read public bug databases on their own to pick up new tricks, then came back and used them. The gym agent wasn't broken either. It did the job it was handed, and nobody told it where to stop. A harness will keep a model going until it hits a limit, and somebody has to set that limit. If you're giving an agent your logins, decide first what it's allowed to touch.

QUICK TAKES

Free Models Had the Biggest Week

The story: Alibaba said its free Qwen models passed 3 billion downloads in six months, ahead of Google's 418 million and Meta's 227 million for all of 2026. Meta gave away Muse Glimmer, a 30 billion parameter model squeezed down to about 17GB so it runs on one consumer graphics card. China's Z.ai announced GLM-5.3 and said the weights ship in roughly two weeks.

Your takeaway: Open weights and open source get used like they mean the same thing. They don't. Open weights means you can download the finished model and run it on your own machine. You don't get the training code or the details of what it learned from. Check the license before you build anything on one, because some let you sell what you make and some don't.

Two Stories About What AI Is Doing to Work

The story: Gen Z trade school signups are up 1,421%, with enrollment at trade-focused two-year colleges up about 20% since spring 2020. Bloomberg found thousands of workers across India getting paid to film their own jobs with phones strapped to their heads, including a 43-year-old who sorts plastic waste in New Delhi for about $211 a month and had no idea what the footage was for.

Your takeaway: A company doesn't need AI that beats every person in the building. It needs AI that gets close enough on one task for less money, and new graduates have worked that out. Hands are still the hard part, so the video that teaches a machine to sort or weld tends to come from the person doing that job today.

Everyone Started Labeling AI, and the Labels Already Wash Off

The story: Anthropic put an invisible watermark inside text Claude writes, to meet an EU rule that started August 2. Spotify said it will badge AI Persona artist profiles in mid-September and keep them out of recommendations. Google went the other direction and let people switch off the visible sparkle logo on Gemini images, videos, and songs, though the hidden SynthID mark stays.

Your takeaway: Anthropic says a heavy rewrite or a translation can strip its own watermark out, and people are already sharing workarounds. Labels are showing up faster than anything that can enforce them.

TOOL OF THE WEEK

🐧 LocalSend Free and Open Source

Twenty-three tools ran this week. This is the one I'd hand to anybody. LocalSend moves files between a Mac, a Windows PC, an Android phone, and an iPhone over your own wi-fi. No account, no upload, no cloud in the middle. If you've ever emailed yourself a photo because the two devices in your hands refused to talk, that's the whole problem it fixes.

Runner-up: BentoPDF, which merges, splits, and compresses PDFs inside your browser. Same idea, different job. Your file never reaches a server.

TRENDING

Mark Zuckerberg promised everyone a free personal AI agent - A 6,500-word letter pitched an agent that tracks your sleep, your money, and your kid's homework, free for billions of people, with a $1 billion fund for towns near Meta's data centers. About 98% of Meta's money comes from ads, and it pulled in roughly $61 billion last quarter. Four days before the letter, a New Mexico judge ordered Meta to pay $942 million over child safety.

Elon Musk's xAI put always-on agents inside your work apps - Grok Bot signs into tools you already use and clicks and types like a person, so it works with services that have no API at all, and it keeps running after you close your laptop. Watch it do a task once and it saves the steps. Pricing starts at $120 a seat and runs to $300.

Apple trained its own model, and only China gets it - After years of paying other companies to run AI on iPhones, Apple built a large language model for the China market with Alibaba's help. Neither ChatGPT nor Claude is allowed there, so renting wasn't an option. Apple's phone sales in China have slipped while local brands shipped handsets with AI already inside.

An unreleased Claude spent 36 hours on a 150-year-old math problem - It didn't solve the Riemann hypothesis, but it pushed a related lower bound from 41.6% to 67.2%, coordinating about 60 sub-agents and burning 31 million output tokens. Two in-house mathematicians checked the result. The prompt came from a staffer with no serious math background who told it to take a real stab and then went for a jog.

She lost her voice to cancer, then got it back from three minutes of audio - Days before surgery took her voice box, Harty, 39, had family dig up about three minutes of old recordings. A speech pathologist at Boston Children's Hospital cleaned up the audio and had a working copy of her voice in under a minute. She types now, and what comes out sounds like her.

Sign language reading ships on the Pixel 11 this Thursday - Google DeepMind's SL2T model turns American Sign Language into English text in real time, arriving August 20 inside Gboard and Live Transcribe. Your phone turns your movements into dots and lines before anything leaves the device, so the raw video stays with you.

PROMPT OF THE WEEK (copy and paste into Claude, ChatGPT, or Gemini)

⚙️ Describe the task you keep rebuilding from scratch. Get a step-by-step system so you stop reinventing it every time.

Six prompts ran this week. Wednesday's wins because it's the one that keeps paying you back. The others answer a question once. This one hands you something you'll reuse every month.

What this does: Describe a task you redo constantly, how often it comes up, who's involved, and where it usually goes wrong. It builds a named, reusable system: a fixed checklist of the steps that never change, decision points for the parts that branch, an optional copy-paste template if the task involves writing something, and the common failure points with a fix for each. A closing line names roughly how much time the system saves once it's running. Everything saves to local storage.

WHERE WE STAND (based on this week's news)

✅ AI Can Now: Run a four-day break-in across 21 connected government systems, choose its own next targets, and catch and fix its own mistakes with nobody stepping in.

❌ Still Can't: Tell your instructions apart from instructions hidden inside a file it's reading. White 3-point text in a PDF or a court filing still works on it.

✅ AI Can Now: Rebuild a person's speaking voice from roughly three minutes of old recordings, in under a minute.

❌ Still Can't: Keep a watermark attached to what it writes. A heavy rewrite or a translation strips the mark out, and Anthropic says so itself.

✅ AI Can Now: Run a 30 billion parameter model on one consumer graphics card, after being squeezed from more than 55GB down to about 17GB.

❌ Still Can't: Beat the top closed models at coding. GLM-5.3 scored 34.5% on Z.ai's own test. Claude Fable 5 scored 39.5% on the same one.