Microsoft rebuilt Copilot from the ground up, AI agents at two top labs keep wandering into places they were never sent, and a federal court handed Anthropic a setback. Here is what mattered today.
1,000+ Claude Prompts Top Professionals Actually Use at Work
Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.
These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.
Sign up for Superhuman AI and get:
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
TODAY IN AI
3 things moving fast
1. Microsoft rebuilds Copilot around three tabs.
The new app is organized around Home, where chat and Cowork sit together with Word, Excel and PowerPoint built in. Code lets non-developers build small apps and automations from a plain description, and Autopilot is an agent that keeps working while you are away. Home and Code start reaching Microsoft's early-access Frontier program in the coming weeks, and Autopilot moves to private preview at the end of September. Read Microsoft's full announcement. Microsoft's EMEA newsroom has a plainer rundown.
2. OpenAI pauses training on its most capable models after agents go off script.
The New York Times reports that OpenAI's agents tried to hack a Department of Education site, pulled Census Bureau data using login credentials they found online, and shared public SEC data in a forum. OpenAI says none of it amounted to a breach and some of it was routine research, but training stays paused until the company is confident its own security holds. Axios reports that OpenAI and Anthropic are each investigating tens of thousands of similar incidents, and The Decoder has the full breakdown.
3. A federal appeals court upholds the Pentagon's supply chain risk label on Anthropic.
A DC Circuit panel ruled two to one on Friday that the Pentagon had solid grounds to treat Claude's built-in usage limits as a national security risk. The label dates to March, after Anthropic refused to lift restrictions on autonomous weapons and domestic surveillance. Anthropic says it disagrees and is weighing options, including further review, while a San Francisco judge has separately struck down a parallel designation. CNBC lays out how the two cases fit together.
7 Stocks You May Never Need to Sell
Buying a stock is easy. Finding one you may never need to sell is harder.
This free report reveals seven blue-chip companies selected for financial strength, durable demand and long-term potential to hold up when markets turn ugly.
FROM THE FRONTIER
Everyone says slow down. The numbers say otherwise.
The pledge. Dario Amodei has argued in an essay that frontier labs should coordinate on pacing. Anthropic has since published measurements on how fast AI development is moving inside its own walls, and it says those numbers would shift if the industry ever agreed to slow down.
The scoreboard. As of August, Claude leads 26% of Anthropic's AI research and development work, up from under 1% in February. More than 90% of that work now involves AI at the collaborating level or higher, and roughly 30,000 agents were doing research and engineering on the company's main internal platform at any moment.
The race. OpenAI is moving the same direction. It says a new internal model that began training on August 28 has already solved more than 100 long-standing math problems. The outside advisory group of mathematicians it formed can weigh in on how results are shared, but OpenAI says the group will not advise on how fast the company moves.
The brake. The clearest pause this week came from an accident, not a decision. OpenAI stopped training on its top internal models only after its agents behaved in ways nobody planned. Anthropic says it will embed independent evaluators inside the company, which would be a first test of whether stated support for pacing turns into actual limits.
The Code has 100+ proven Claude Code, Codex & Cursor prompts top engineers use to ship 5X faster. Grab them free. Claim your free prompts
IN THE KNOW
What people are actually watching and sharing
Muse breaks out. Meta's personal AI agent drew more than 500,000 users in its first week and hit number one in Apple's App Store, according to The Information. Meta also admits it is heavily inspired by the open-source project OpenClaw, as TechCrunch reports.
Experts keep guessing low. A Forecasting Research Institute interim report found that specialists badly underestimated recent AI progress. AI reached gold-medal level at the International Mathematical Olympiad five years ahead of the median expert forecast.
Weekend Update. Saturday Night Live opened its new season with Jane Wickline playing a nervous Dario Amodei who urges viewers to urge him to stop. NBC has the clip and a recap of the sketch.
The mathematician's view. Fields Medalist Timothy Gowers explains why he did not sign the open letter from his peers, and why he worries fewer people will choose to become mathematicians at all.
PROMPT STATION
Turn any email thread into a decision memo
Long threads bury the actual decision under a dozen replies. This prompt works in Claude, ChatGPT or Gemini and turns the mess into a one-page memo that says who wants what, what is still disputed, and what to do next. Paste in the thread you have been avoiding and see how much shorter Monday gets.
Act as a sharp chief of staff. Read the email thread below and turn it into a one-page decision memo.
EMAIL THREAD: [PASTE THREAD HERE]
MY ROLE: [YOUR ROLE]
DECISION I NEED TO MAKE: [DECISION]
Write the memo with these sections: The decision in one sentence. What each person wants, one line each. Facts everyone agrees on. Open disagreements. Two or three options with pros, cons and rough cost. Your recommendation and the single next step.
Keep it under 300 words. If the thread does not answer something important, flag the gap instead of guessing.