Claude Duet Works Between Chats; Math Proof, Jobs Data and PHP Flaws

Welcome to the Superpower Daily Briefing for October 22nd 2026 I'm your host and we've got a fantastic whiteboard explainer lined up for you today mapping out the absolute biggest developments in AI Let's jump right in We are tracking four massive stories today First up Anthropic's new persistent AI Then how human guided AI just cracked a 31 year old math problem After that we're looking at AI's impact on young workers And finally how AI helped uncover some

incredibly critical PHP server flaws Let's get to it All right Chapter 1 Anthropic just opened a wait list for Claude Duet This is a brand new research preview that basically gives your AI a job it can just keep doing even after you completely close out the chat window Now this perfectly highlights the massive shift we're seeing in AI right now You know how regular Claude just sits there waiting for you to prompt it before it does

anything Well Duet doesn't wait It actually runs in the background using its very own cloud computer and browser It's out there checking your connected accounts and preparing work even if your laptop is literally shut in your bag To give you a super concrete example of how this plays out imagine Duet automatically drafting a meeting brief based on your connected apps Then if it spots a conflict say with an upcoming flight you booked it'll proactively reach out

and offer to reschedule the meeting You don't even have to ask But what's really fascinating here are the firm boundaries Anthropic put in place Yes this transforms the assistant into a persistent agent It reads it drafts it preps work totally autonomously But and this is a big but it absolutely must wait for your explicit approval before it actually sends any messages or signs documents It has to ask before making a payment with tools like link or

shop pay and it completely cannot trade securities or move money around your accounts No rogue day trading allowed There is one major catch though and it comes down to resources Duet's background checks actually drain your existing five hour session and weekly usage quotas So while having an autonomous agent working in the background is incredibly useful it comes with a very real cost to your standard chat limits All right let's move on to chapter two and see

how this whole theme of human and AI interaction is evolving Microsoft researchers Xinjie Zhang and Jiao Song have actually managed to remove the dimension dependence from a 31 year old math puzzle and they did it using human directed AI Okay so the puzzle they solved is known as the KLS conjecture To translate some incredibly complex geometry into plain English it's basically a theory predicting that a simple flat cut can capture a shape's narrowest bottleneck almost exactly

as well as some complex twisting cut could And the huge breakthrough here they proved this holds true no matter how many dimensions you throw at it Now the exact workflow they used is just a textbook example of where modern research is heading This wasn't some autonomous AI magically solving the problem while the researchers went out for coffee Humans set the mathematical direction first then AI agents rapidly explored over 100 candidate approaches in parallel After that the

humans stepped right back in to pick the most promising results piece together the actual argument and then rigorously check the final proof So why does this matter to the rest of us Well a lot of foundational math and algorithm research out there kind of just had to assume the KLS conjecture was true to keep moving forward Now they can proceed by stating it as absolute fact no assumptions needed It's a huge foundation completely solidified But there's

a vital limitation to keep in mind and the researchers pointed this out themselves They noted that several leading AI models entirely fell short even when given roughly the same total token budget Unguided fully autonomous models failed completely here So this is purely a story of human directed success The humans provided the absolutely crucial judgment Moving right along to chapter three let's pivot over to the labor market Stanford's Canaries dashboard is currently throwing up a red flag

about a prolonged employment drop specifically targeting young workers in highly AI exposed jobs Using totally anonymized ADP payroll records their dashboard tracked a 4.4% year over year employment decline this past August specifically for workers aged 22 to 25 To put that into perspective this drop is heavily impacting highly exposed fields like software development and customer service But you know what's really striking Employment in these specific segments has contracted every single month since October of 2023 Now

you might be thinking why are we tracking such narrow slices of data Well monitoring these specific segments is absolutely crucial because the dashboard essentially acts as an early warning system It's custom built to catch those emerging underlying labor trends that massive economy wide employment totals usually just smooth right over But and this is super important the big unanswered question here is causation This data firmly shows exposure and correlation but it doesn't definitively prove that AI actually

caused these specific job losses Plus while the dashboard did hint at lagging job growth for young women earlier those gender gap theories remain unproven Longer term government statistics just aren't showing hard evidence of a widening gap quite yet Finally for chapter four we are diving into cybersecurity PHP just had to release a critical patch after researchers at Wiz basically used AI to turn some long standing memory bugs into full blown server takeover attacks Let's look at

the timeline on this because it's pretty wild Over ten years ago some vulnerable fatal error handling code was introduced to PHP Ten years Fast forward to today and researchers using frontier AI models managed to compress the incredibly complex process of refining these exploits down to just a matter of hours So on October 21st 2026 PHP had to rush to get a critical patch out the door To show you exactly how dangerous this is the researchers demonstrated

what's called remote code execution Basically that means forcing a server to run an attacker's malicious instructions And they did this on WordPress Drupal and Joomla deployments without even needing to log in That kind of access could instantly expose sensitive passwords and critical OPI keys And here is why this matters on a massive macro scale The Wall Street Journal estimates that PHP powers up to 70% of websites globally Now to be fair that's not a direct count

of vulnerable websites But it really underscores why AI accelerated exploit development is becoming such a massive structural risk for the whole industry right now So until everything is fully updated defenders can keep an eye out for warning signs Things like malformed multi part requests oversized session cookies or just repeated HTTP 500 server errors But make no mistake the ultimate takeaway here is that the only complete remediation to actually secure these servers is updating to that newly

patched PHP version All right let's wrap up today's explainer with four really concise takeaways from the stories we covered today Number one Cloud Duet is bringing us incredible persistent background work But remember it will drain your existing usage quotas Number two AI is definitely speeding up complex math proofs But right now it still absolutely requires strict human guidance to succeed Number three Stanford's data is signaling a real job contraction for young workers in highly AI exposed

roles even though the exact causation isn't fully proven yet And number four AI's emerging ability to rapidly weaponize decades old server flaws makes immediate patching of systems like PHP more critical than it has ever been which leaves us with one final question to ponder as we head out today How will the evolving balance between autonomous AI execution and necessary human oversight end up shaping your daily work Read more at superpoweredaily com Stay curious