Claude can keep working while your laptop is off

Welcome to The Signal from Superpower Daily with Maya and Theo

Today Anthropic is giving Claude a job it can keep doing after you close the chat

Yeah

imagine you finish your workday

You

know

You

close your laptop

you put it in your bag

and you just go to sleep

Right

And while you are dreaming

your AI assistant is reading your incoming emails

It is checking your bank accounts

It is even drafting client briefs for the next morning

Which is wild

And that is exactly what we are looking at today

So Anthropic just opened a research preview for an assistant they call Claude Duet

And the core idea here is just practical

This AI actually continues working between your chats

Right

It

does not freeze the moment you close the browser tab Right

I

mean usually you close the tab and the AI is just gone But

Anthropic launched this as a waitlist specifically for their U S subscribers

So this is on their Claude Pro and Claude Max tiers

Yeah and it runs on their new Claude Opus 5 5

model But the mechanism behind this is what makes it so

distinct How

so Well Duet utilizes its own cloud computer and

browser Oh

interesting So it is not

local Right It is not running in the background of your phone draining your

battery It is spinning up a virtual environment on Anthropic's actual

servers

OK that makes sense

And from there it monitors your connected

accounts You can link things like your

calendar shop

pay or specific workflow

apps

I see

And Duet just checks them every few hours while you are completely

offline So it prepares work autonomously We are talking about it drafting replies to threads or creating meeting briefs or even generating

financial

forecasts Yes And you do not have to sit there and watch a

loading bar You just wake up and the work is basically waiting in

your

queue Exactly I mean this represents a massive shift in how we interact with large

language models It

really does For the last few years the paradigm has been

strictly turn based You

prompt the

model It

replies You

wait

It waits It is a highly capable but completely

reactive chatbot Yes But Claude Duet introduces a structural shift toward a persistent

digital worker The real differentiator here is

unsolicited preparation

Unsolicited preparation Anticipate

the

assistant

Oh so it anticipates what you will need before you even ask for

it

Exactly The assistant anticipates what you will need before you actually formulate a

request Let me give a concrete example of this from the testing

phase Anthropic highlighted a brand strategist named Maya

Dalton She noted how Duet proactively prepared a brief for an upcoming client

call Which is helpful on its

own Right But it went further than just reading her

calendar It noticed that the scheduled call fell shortly before a flight she had

booked

Wow Yeah And Duet proactively offered

to reschedule the call for

her It connected the dots across her

calendar her travel

apps and her task list entirely on its

own And that happens because Duet basically inherits your account

memory Oh

right It has your skills

profile It has your connected

apps So if you send a quick voice memo through an ongoing

thread Duet already has the entire context of your

week That is incredibly

powerful It

is And Anthropic actually built a completely separate interface for

this There

is a dedicated tasks tab Okay

This

tab separates these ongoing background jobs from your standard conversational chats And

it highlights work that needs your human input Because

only you can mark a task as completely finished Wait

I

need to push back on this a bit Because you know autonomy sounds great until you look at the

cost There is a massive limitation here regarding your usage

quota Oh

absolutely Background work draws directly from your existing five hour and weekly limits

on Claude That is the defining bottleneck

right now There is no separate pool of usage for

this preview None

at

all Nope There is no extra

fee either to

be fair But because

it draws from your standard

allowance every single morning brief it generates counts against your

limits

Yeah Every time it wakes up to check your

inbox it burns

tokens So it could just drain your

account

Well Anthropic spaced these checks out to every few hours rather than monitoring

continuously And they did that specifically to contain that token

consumption

Honestly it is like hiring a highly eager

intern You give them your corporate expense

account And before you even step into the office they have spent half your budget buying research reports you might not even

need That is the perfect

analogy And that is the exact tension users will

face You have this incredible desire for autonomous

help but it clashes violently with quota

anxiety

Quota anxiety

Imagine you are a power user

You rely on Claude for intense coding or complex data

analysis during your actual workday Sure

The

last thing you want is to log in at 10 a m ready to work only to find out your assistant burned through half your weekly tokens organizing your spam folder at four in the

morning

Exactly So how do you actually control an eager system like that

Anthropic built in some very strict boundaries

right

They had to

So Duet can read draft

and prepare all on its own

but it requires manual user approval before taking any consequential actions

Consequential actions

like what

If it wants to send an email

make a payment

sign a document

or delete files

it has to stop

It just waits for you to say yes

And if you are not there

If you are not around to approve it

it just parks that task and moves on to something else

The payment controls specifically reveal how carefully they are treading here

Purchases use Stripe's Link Wallet or Shopify's ShopPay

Right

And these are connected directly inside Duet

This is completely separate from the normal Stripe connectors a business might use for analytics

Yeah

let me explain how that actually functions when a payment is requested

Please do

If Duet wants to buy say

a software subscription using Link

it pings you

You must approve the exact amount in the Link app on your phone

Okay

so a hard physical approval

Once

you approve it Link

generates a one time virtual card

That card is limited specifically to that single merchant and that exact dollar amount

That is pretty smart

ShopPay works the exact same way

Plus Anthropic hard coded some absolute

limits Duet cannot trade

securities It cannot move money between your bank

accounts Even if I tell it

to Even if you explicitly try to give it

permission it just will not do

it There is a massive security angle here

too

though Anthropic keeps your passwords and sign in tokens completely outside the model's

direct

view Yes that

is crucial They use a screening system to block it from sharing credit card

numbers or your

government ID but they also explicitly warn about

prompt injection Prompt injection is a critical vulnerability for

background agents Walk us

through that Imagine an

incoming email That email could contain hidden malicious

text instructions If Duet reads that email in

the background those instructions might attempt to hijack

the agent So the AI gets confused by the

hidden

text Right The malicious text could command the AI to forward sensitive documents or even delete

your inbox Which is exactly why that human approval choke point

is mandatory But you know that creates a huge

responsiveness issue

How

so Well it only checks in every

few hours It is

not continuous That means Duet might completely miss a highly urgent message that arrives right after its

last check

Oh

true Furthermore during this

preview phase

you cannot watch its browser

You cannot take over the screen

You just have to trust the results it brings back to you on the tasks tab

And that requires a massive leap of faith for professional workflow

If you disconnect an app

it stops future access

but it does not erase the memory of what it already saw

Wait really Yeah

not

unless you completely wipe your chat history It

is also missing major integrations right now like Microsoft 365 is not supported yet So

if you were watching this space the

next thing to monitor is user behavior Will

professionals actually trust unseen background work enough to justify burning their standard cloud usage allowances It

is a big question Are

those unsolicited drafts valuable enough to warrant the token cost Or

will quota anxiety turn this into a novelty you just turn

off to save tokens for your actual workday It

is the ultimate test of AI utility If

the background work is flawless you

will happily pay the token cost If

it hallucinates or if it prioritizes the wrong tasks while you are asleep you

will disable it immediately Yeah

totally Well

that wraps our focus on Claude Duet's autonomous background work While

Anthropic is pushing for AI to work entirely out of sight our

next story looks at what happens when humans keep AI on an incredibly tight leash Microsoft

researchers just detailed how human guided AI helped solve a 31 year old

math problem And this story really shatters the romanticized version of

scientific

AI

Oh completely The popular myth is that you feed a massive problem into

a

model go get a

coffee and return to a Nobel Prize winning

discovery

Right the AI just does it

all But the reality is incredibly labor intensive for the human

scientists So here's what actually

happened Microsoft researchers Xinjie Zhang and Zhou Song used AI agents to prove something called the Kanan

Lovasz Simonovitz conjecture Usually

just called

KLS Right KLS It is a 31 year old geometry problem that underpins

the study of randomized algorithms And they directed the AI through more

than 100 candidate approaches simultaneously The

key word there is directed The AI did not just solve the problem from start to finish The

humans acted as strict supervisors So

they were steering the ship Exactly They discarded the bad mathematical paths They took the most promising results from the AI's intermediate work

Then they fed those specific results back into

the

next round of prompting

And by October 4 they achieved a recursive refinement that completely removed the dimension

dependence

from the bound OK I want to break down what KLS actually is because dimension

dependence sounds incredibly dense It

is

a bit abstract yeah So this problem deals with bottlenecks

and high dimensional probability distributions specifically a

class

called

log

concave distributions Right Think of

a bell curve That is a probability distribution

in two

dimensions Now imagine a complex multidimensional

landscape You are trying to find the narrowest passage or the bottleneck within that

space These bottlenecks dictate how quickly a random walk algorithm can explore the entire

distribution Let me try a metaphor

here

Imagine pouring water out of a jug

If

the jug has a wide neck the

water flows out instantly If

it has a narrow bottleneck the flow is restricted Algorithms

exploring data work the exact same way That

is a great way to picture it Thanks

The KLS conjecture basically predicts that a simple flat slice through this mathematical space can locate that narrowest passage

You do not need to calculate a complex wavy cut

A flat cut works

And

crucially KLS

predicts this holds true regardless of how many dimensions you add And

that right there is the dimension dependence issue Back

in July 2026 the best known mathematical proof still grew slowly as dimensions increased

It

meant the guarantee weakened in infinitely complex spaces Zhang

and Song published a paper on October 1st that reduced that dependence to an iterated logarithm Okay

That

grows extraordinarily slowly but it still grows But

then came October 4th Using

a recursive refinement method they

removed the dimension dependence entirely They

proved

that the errors introduced along the way shrink fast enough that the total error remains mathematically bounded So KLS

holds true in infinite dimensions And the

reason this is a top story is not just the geometry It highlights

a realistic modern workflow for elite science How so

The AI

handled the tedious brutal derivations It

crunched the higher order tensor calculations Which are what

exactly Tensors are

essentially massive multidimensional grids of numbers They are punishing

for a human brain to track but trivial for

a machine However the humans

provided

the mathematical criteria The humans assembled the final logical proof But you know

if the AI generated over 100 distinct approaches and the AI crunched the massive tensor calculations where exactly is

the line What do you

mean Is this a

human solving a problem with a tool Or is this

a tool solving a problem with a human supervisor Well the researchers

stated that the key mathematical criterion came from them The spark of

insight that allowed their recursive method to work was human Generating 100 approaches

is

just throwing mathematical spaghetti at the wall The human knew which approach

was actually elegant and mathematically sound So the AI provides the

sheer scale of exploration It instantly derives the consequences

of 100 variations A human might spend a lifetime

going down just three of those blind alleys But without the human's aesthetic judgment

to

recognize the correct path

the AI just generates endless useless derivations

There is an important caveat

in how

this research is being presented though In their account the researchers noted they

tried giving the

problem to leading AI models in an unguided way Yes they did They gave the autonomous

models roughly the

same total token budget And they report that the unguided models

failed completely KLS was not solved by an AI

just grinding away on its own The limitation here is that this is

strictly a historical narrative from the researchers It is their personal account of their

project workflow It is not a benchmark Right It

is not a formal independent

benchmark of model capabilities We only have their word that unguided

models failed on that specific prompt budget Still the outcome for

computer science is massive A

separate paper by Pierre Bazil Boaz Klartag and

Joseph Lehek quickly confirmed that Zhang

and Song's criterion was the crucial step They built on it KLS is resolved That is

huge That means countless

algorithms that simply

assumed KLS was

true can now operate with absolute mathematical certainty And if you are following this field you need

to watch for the researchers to release

their shortened proof More importantly watch for its machine checked Lean 4

formalization after

final peer review Let's explain Lean 4 for a second Sure

Lean 4 is a programming language specifically

designed

for theorem proving Getting a proof written in Lean is the

modern gold standard for mathematical verification Because it is so rigorous Exactly It forces

the logic It'd be so

explicit

that a separate computer program can

verify

every single step is flawless It removes human trust from the equation entirely I see You

use humans and AI to find the proof

you use

a strict computational tool to prove you did not

hallucinate the logic Which is fascinating But you know if elite AI is acting

as an ultra

fast math assistant for human geniuses it raises a very uncomfortable question What happens to the human assistants That is a very

real concern Our next top story

looks exactly at that Stanford is

tracking a 4 4 employment drop for young

workers in AI exposed jobs Yeah we have spent years debating the theoretical

anxiety around generative AI and white collar work This brings that theoretical fear into hard

empirical data It really does The Stanford Digital Economy

Lab partnered with

ADP Research to build the Canaries Dashboard And the data it is outputting is

sobering Here is what the dashboard revealed U

S employment for 22 to 25

year olds in highly AI exposed roles dropped 4 4 year over year in August That is

a specific demographic We are

talking about entry level knowledge work software development customer

service

financial analysis

These are

the exact kinds of roles you take right out of college to get your foot in the door France 24

initially reported this August data point on October 6th Stanford then published their detailed methodological account on October 21st And the

reason this matters deeply is the timeline right Exactly

This

is

not a sudden October shock It is

not a one off blip in the data This

is a sustained pattern occurring every single month since October 2023 We

are looking at a prolonged contraction in entry level white collar

work And the methodology here is what makes it so

credible Broad economy wide government statistics often smooth over specific

demographic hits The Bureau of Labor Statistics relies heavily

on surveys But Stanford and ADP are using anonymized actual

payroll records Right because surveys have massive lag times and rely on

human memory Payroll records are literal dollars hitting

bank accounts You cannot fix a

payroll

record By grouping those payroll deposits strictly by occupational AI exposure the researchers isolated a very narrow labor market segment So

if you are 23 years old and your job consists of tasks an AI can easily do there

are 4 4 fewer of you on payrolls this year than last year

Yes They call it the canaries dashboard We are seeing the canary coughing But we have to highlight a massive

flashing neon caveat here

Which is

Correlation is not causation

We do not actually have proof that AI is the gas in the coal mine

That caveat is essential

The data tracks employment in exposed roles

It does not prove the algorithm directly caused the job losses

Right

There are other factors

Oh absolutely

We have had high interest rates battering the tech sector for over a year

Venture capital dried up

Startups pulled back on hiring junior developers

just to extend their runways And

customer service centers might be offshoring those roles to different countries for macroeconomic reasons That

happens completely independently of AI Exactly

The

dashboard shows us precisely where the pain is happening It

does not definitively isolate the algorithm as the executioner That

is a crucial distinction There

is also a secondary signal in this dashboard that requires extreme caution The

data hints at slower job growth for young women compared to young men in these exposed roles Stanford

points out that other research suggests women disproportionately work in these highly exposed fields But

the researchers themselves flag a problem with this conclusion The

data

representation ADP payroll

data might not be representative enough of the total U S population to confirm a gender gap Long

term government stats are not showing this specific gender split yet

It is an unresolved signal

So if you are an operator or an investor

you need to watch if broader government employment data eventually confirms this early warning signal

If the Bureau of Labor Statistics starts showing structural rot in entry level knowledge

work across the entire economy

then the coughing canary was right

And the downstream effects of this are terrifying

If you eliminate the entry level jobs where junior developers learn how to code

securely you create a massive talent pipeline crisis five years from

now You really

do You cannot just spawn senior talent out of thin

air That structural pressure on entry level human jobs

is intense But there is a different kind of structural pressure hitting the actual code that runs

the Internet Which brings us to our third

top

story PHP has patched critical flaws after AI assisted research revealed severe server takeover

risks This story is a masterclass how

AI is fundamentally altering the timeline of cyber security It is accelerating both the attackers and the defenders So

on October 21 PHP

released patches for critical bugs Specifically

these are tracked as CVE 2026 109618 and

109619 And these vulnerabilities carry a CVSS severity score of 9 2 out

of 10

Wow 9 2 That is

alarm bell territory A 9 2 means these

flaws allowed unauthenticated

remote code execution Let's explain what that means in plain English

for a second Unauthenticated means the attacker does not

need a password They do not

need an account

None at all Remote means they can do it from

anywhere on earth And code execution means they can run any command they want

on your server They completely

own the machine And this impacts massive platforms built on

PHP like WordPress

Drupal and Joomla But the real kicker is how this

vulnerability was weaponized This is

the crazy part A security firm called Wiz used Frontier AI models to build and adapt working exploits

in just hours They specifically

used Wiz Atlas

Gemini 4 Argon and

GPT 5 6 Sol To understand why this matters beyond the headline we really have to look

at the vulnerability itself This is not a sloppy piece of code in

a random WordPress plugin The vulnerability lives deep within the

architecture

of PHP itself It stems from how PHP handles fatal errors

and memory exhaustion So when a server request runs

out of memory PHP can leave behind a

corrupted internal state Think of a

bustling restaurant kitchen Okay

I'm with you A chef is preparing

a complex dish Suddenly they run out of

a crucial ingredient That

is memory exhaustion Instead of cleaning

up their station they

just walk away leaving a half prepped corrupted

dish on the counter Then the next chef comes to that station to prep

a completely different order They unknowingly use

that

corrupted half

prepped dish In PHP if later requests reuse

that corrupted memory state an

attacker can manipulate it They can

read sensitive process memory they can steal authentication

tokens and API keys or they can just trick the server

into executing malicious code Wiz

called their disclosure PHP to shell They pointed out that this specific vulnerability

dates back

to PHP version 5 4 Version 5 4 Yeah That means this ticking time bomb has been sitting in the underlying

architecture of the web for over a decade Even a basic Hello World app running

an affected PHP version could theoretically be exploited But the real reason this

is a top story is the AI angle Adapting an

exploit like this requires manipulating deep memory allocation It requires predicting how

memory is reused across completely different

server setups And

that used to take

immense

painstaking manual effort It was artisan work Right A human security

researcher would have to manually link the bugs They would have to account for the

unique memory behaviors of a

specific Linux deployment

They would run the attack analyze the

failure revise the

payload and

test it again It took

months But now Problem to Gemini 4 0 Argon and GPT 5 6 Sol They reduced that artisan

labor to practically zero The models analyzed the

failures They debugged the attack

paths They adapted the exploit

for different operating systems in hours It is incredibly ironic

really Very much so

We

constantly hear that

AI is a defender's ultimate shield Yet here it successfully

weaponized a decade old bug across multiple frameworks faster than any human could It is the