if this doesn't scare you, you're not paying attention. "after a model exploited a dns misconfiguration to find a way outside of this offline world, we paused frontier training" - OpenAI my agents run all night behind one claude.md. would you let yours?

Sam Altman: "i realized i have not read or even looked at a line of code in probably a year." i'm not far behind: 5 grok bots and opus 5.5 write mine while i sleep, and a reviewer checks every diff at 7. setup below:

Dario Amodei after finding out an open-weight model now claims to match the Opus 5.5 at coding and agents

claude code just reverse-engineered a WIFI router and found why it was slow in 5 minutes now imagine it working all night while you sleep. one settings file makes that safe:

Google CEO, Sundar Pichai: "We are pretty far from the end state of RSI" ai that improves itself is just getting started. the devs pulling ahead right now all set up one file first:

CLAUDE CAN NOW SPEND YOUR MONEY stripe just plugged into claude duet, the proactive agent anthropic opened yesterday -> with link it can order product samples, send gifts or top up office supplies -> with the stripe MCP it can issue refunds and pull your business trends the crazy part is the guardrail. you approve every single purchase, and your payment details stay private to be fair duet is still a waitlist research preview, pro and max in the US only, so most people won't touch this for a while personally refunds are the part i'd try first. a customer asks, duet finds the order in stripe, asks me once and it's done

most people set up grok bot wrong and it's f*cking painful to watch... they build 5 bots that write and zero that check. by day three the bots go quiet and nobody can tell which one broke the reviewer goes first. it never writes, it only says SHIP or SEND BACK. all 6 bots and their full prompts are in the article below

Elon Musk on Grok Bot: "We're seeing the fastest growth of a product that I've ever seen in my life" he says people are using it more every week, not less, and enterprise grok bot will be in wide deployment in a few weeks every company that turns it on is going to need someone who knows how to set it up. the article below shows 5 setups that already have buyers and how to sell them

ANTHROPIC JUST KILLED OPENAI'S DOTS anthropic just opened claude duet, a new app where claude works as a proactive agent -> it comes to you with what needs your attention -> it finishes the tasks you hand it -> it keeps working when you walk away -> it checks with you before it acts on your behalf it runs on opus 5.5 and pulls in your memory, skills and connected apps from your claude account, so there's almost nothing to set up the crazy part is the safety layer. it never directly sees your passwords, and a new classifier guards against it sharing anything sensitive to be fair dots is live and duet is still a waitlist research preview, pro and max in the US only. anthropic even says it can still misjudge what's urgent personally i'm joining the waitlist tonight. the first thing i want to see is what it decides needs me meanwhile at openai HQ:

AI FOUND A HOLE IN 70% OF THE WEB openai says its models helped researchers find critical security flaws in PHP, the language behind an estimated 70% of websites -> 2 CVEs, rated 9.2 out of 10 by PHP -> remote code execution, no login needed on mod_php -> one bug can leak passwords, tokens and API keys from other requests the crazy part is attackers were already trying it before the patch shipped. patchstack and cloudflare saw the attempts openai spent two months on this with a coalition of maintainers, researchers and security providers. fixes went out today: PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to be fair the models helped researchers find it, people still had to confirm every bug and write the fix. and if you're on PHP 8.1 or older there's no official patch at all personally this is the AI use i want more of. if you run a php site, update today and restart php, a package upgrade alone keeps the old version running

CHATGPT CRACKED A BRAIN DISEASE CLUE IN 32 MINUTES openai just showed what gpt-rosalind did with scripps research and @perturbaitx thousands of agents ran in parallel, matched gene data from millions of mouse brain cells against the scientific literature, and came back in 32 minutes with a new lead for a brain disease -> a gene linked to perry syndrome, a rare neurodegenerative disease, switches on cholesterol-making genes when it's disrupted -> rosalind's guess: the cholesterol isn't reaching where neurons need it -> if that holds, fixing cholesterol transport becomes a direction for new treatments the crazy part is the second find from the same data. TSC1 and TSC2 work as a pair, but in one group of brain neurons, breaking TSC2 changed dozens of genes and breaking TSC1 barely did anything. that could explain why TSC2 mutations hit harder to be fair these are leads, not cures. openai's own words are "a specific mechanism to test", and it still has to survive the lab personally the speed is what gets me. a question like this now comes back with a list of things to test in half an hour

this grok bot setup is basically a content agency in one prompt and it's f*cking insane... 4 bots: one scouts what's working this week, one pulls the numbers, one writes, and one throws out every line that has no source clients already pay writers $2K-$5K a month for this the full build and how to sell it is in the article below

Meta's AI chief, Alexandr Wang: "we actually delayed the launch of Muse for many months to get the safety and security exactly right" in 68 minutes he explains what it takes to trust an AI agent with your personal info, your phone calls and your tasks, and why meta built a separate sentinel agent just to watch what muse sends out your own agent needs the same thing on a smaller scale. the article below shows the one file that keeps it in line while you sleep

this prompt finds you clients while you sleep and it's f*cking insane... drop it into grok bot tonight. by 8am you get 10 local businesses that need an AI setup, the proof they need it and the first message already written it never contacts anyone on its own. you wake up and pick the prompt is in the picture. how to price and sell the setup is in the article below

EVERY KID IN AMERICA GETS SPACEX yesterday anthropic said it is putting 5% of its shares into a new public benefit fund, plus up to 2% more every time the company's value grows 50% over a baseline add the founders pledging 80% of their equity and employees donating theirs, and roughly 20% of anthropic is now committed to the public good two hours later elon replied "I recommend giving the shares directly to the public" this morning he went further. one of his spacex shares to every kid in america next year, and he says "this will be the first of many" the crazy part is who is doing this. the companies building the AI everyone is scared of losing their job to are the ones starting to hand out pieces of themselves to be fair nobody knows yet what the anthropic fund will actually pay for, and one share per kid is a nice gift that wont pay anyone's rent personally i want to see where the fund's money goes first. if it ends up paying people who lose work to AI, every other lab is going to get asked to match it