AI finds cancer survival clues in face photos

Welcome to The Signal from Superpower Daily with Maya and Theo

Hey everyone

If you took a selfie today

what are the chances an algorithm could look at the bridge of your nose and predict whether you will be alive in 90 days

Well

right now

the chances are actually pretty good

Because researchers just proved it is not just possible it

is happening right now It

really is Our

mission in this deep dive is to give you a high signal information

dense download of today's most crucial AI developments

Exactly

We are filtering out the noise to bring you the insight

without the information overload

So let us get right into it It is a remarkable shift in medical imaging

Yeah it is

We are looking at a study published on October 19th The journal is The Lancet Digital Health

Right And researchers at Mass General Brigham have developed a system called

Fairface And the scale of this research is

staggering like truly

staggering It is

massive They trained this foundation model on more

than 40 million facial

images

Yeah 40

million Think about that number for a second 40 million

faces We are not talking about a small curated medical dataset here

No not at all

This is a massive intake of visual data

And that volume is exactly why this is a foundation model

Right

You see

a foundation model does not start with a specific medical task

It is not looking for a disease right out of the gate

Exactly

It first learns a very general

highly dimensional representation of human faces

It

maps the geometry It maps the textures It

maps the subtle variations in pixel gradients across millions of parameters So

it is just learning what a face is Right

It builds a mathematical understanding of what a face is It

creates a latent space of facial architecture Yes

It

learns the baseline rules of human features before anyone ever asks it to look for a disease Precisely

And once that foundational understanding was established the

researchers adapted it They

gave it a job They

actually branched it into two distinct tools Okay

The

first tool is called FAIR FaceAge Its

dedicated job is to estimate a patient's biological age based

purely on their photograph And

then the second tool is called FAHR FaceSurvival Right And

this

is the one that really stops you in your tracks This is

quite sobering Its job

is to predict survival risk Yes It

looks at a photograph of a cancer patient and

directly estimates their mortality risk And

they put these tools to the test on some massive patient groups They

really did The

first test group was highly specific It

included 3 820 patients Okay The researchers isolated

the

subgroup that the AI rated at the absolute highest risk And the results were

stark Very stark That specific

high risk

subgroup showed substantial mortality within just 90 days of the photograph being taken 90 days That

is an

incredibly narrow window It is It

is basically

the immediate near term Right

Now we need

to

be clear about the statistics here Please do This specific

subgroup result

does not represent the overall accuracy of the entire model across all 3 820 patients That is an

important distinction But it is

a powerful indicator of what the system can spot in extreme cases Exactly It is

identifying patients who are already in a critical systemic decline And then they

zoomed out They did They

looked at

a separate much larger dataset

How large This one

contained more

than 38 000 patients Okay And they used the

FAR Face

Age tool on this group The age estimator Right They compared

the AI's estimate

of

the patient's biological age against their actual chronological age from their birth certificate They looked

for the gap Yes

They wanted to see what

happens

when you look older than you actually are And they found a definitive link

What was it Patients whose estimated

biological age was

significantly older than their actual chronological age well they had a higher mortality

risk

So the visible acceleration of aging

correlated with poorer survival outcomes Exactly But what's fascinating here is

how

these two different tools interacted Right Because you might assume they're

just measuring the same underlying concept Naturally You might think looking older

is the exact same thing as looking sick Yeah That seems like common sense

But they were not just

giving the same answer in two different ways Wait Really The readings from

FASH

Face

Age and FASH Face Survival were only weakly correlated Only weakly

correlated That is

wild That weak

correlation is arguably the most crucial discovery in the entire study Because if

the two models produced highly correlated results it would

mean they are redundant Exactly It

would

mean they're capturing the exact

same biological signal Just giving you

the same bad news twice Right But a

weak

correlation means they are orthogonal They're capturing complementary

distinct health information They are two

entirely separate puzzle pieces Yes The survival

model

is seeing something that has absolutely nothing to do with traditional aging Exactly And when

the

researchers analyzed the two scores together each reading remained

independently associated with survival So the age

estimate predicted survival And the survival

estimate

predicted survival Yes And they did

not

cancel each other out Combining the age estimate

and the survival risk estimate actually improved the overall accuracy It did It created

a tighter

fit with the actual patient outcomes And we can actually

see how these two models diverge just by looking at their attention mapping Right Attention mapping is

how

researchers visualize the black box of an AI It is like a

window into its brain It is A technique

like gradient weighted class activation mapping That is a

mouthful It is But

basically it highlights which specific pixel regions of an image the model is relying on the most To make its

mathematical prediction Exactly It creates

a

heat map over the face And the heat

maps for the two tools were completely different Completely different For

the biological

age estimates the AI focused

heavily on the lower

half of the face Right It looked at the

creases

between the nose and the mouth The nasolabial folds Yes And

it also looked

closely

at parts of the forehead That aligns perfectly with human

intuition It really does We naturally

associate deep folds and forehead wrinkles with the passage of time The AI learned the exact

same visual shorthand for chronological aging that we use But the survival estimates threw

that intuition out the window Totally out the window When

predicting 90 day mortality

risk the AI ignored the

wrinkles It did not care

about them at all No It focused intensely

on the area directly below the eyes Okay And it

also

focused heavily around the nasal bridge And this is

where the science hits a wall It really is

The researchers do

not know why They have no

idea The mechanism remains

entirely unclear We do not

know the underlying biology connecting the nasal bridge to cancer mortality We do not

We can speculate

about the physiology though Sure Like

90

days

before death a cancer patient

might be experiencing cachexia Right Cachexia is

severe

systemic tissue wasting So we might

be seeing the loss of a very specific subcutaneous fat pad right near the nasal bridge It could also

be vascular How so The

model might

be detecting microvascular changes It might be

seeing subtle shifts in blood oxygenation just below the thin skin of the lower eyelid Wow It could

be detecting subtle fluid retention or microedema So the

AI sees a reliable mathematical pattern but it

cannot

explain the actual biology to us Right Okay Let's

unpack

this

Sure We have

an

algorithm looking at the shape and texture of facial features to predict internal health and imminent mortality We do Does

not assessing

someone's internal health based on the topography of their face border on high tech phrenology You know

that is a very valid historical comparison I mean

phrenology was

the debunked 19th century pseudoscience Yes Practitioners

claimed they could predict mental traits and criminal tendencies by

measuring the bumps on a person's skull Which was completely

arbitrary and deeply biased It was used

to justify terrible things Exactly And the

core

difference here is the empirical rigor Okay The

AI

is not applying an arbitrary theory It is finding statistically significant correlations in massive verified medical data

sets So the mathematical

link is real But your skepticism

is

warranted Thank you It points

exactly to

why we must be incredibly careful about the data we feed these systems Because the data is

the massive limitation of this study It is The survival

model drew

on a cancer patient population from the Boston area Okay And that specific

population

was

majority white And that is a

critical flaw A huge flaw You

cannot deploy a

medical tool globally if it only understands one demographic Think about the underlying

mechanism we just discussed Right If the model

is

detecting mortality by looking at microvascular changes or subtle variations in

pixel gradients on the skin Melanin matters Melanin matters

immensely It

changes how light scatters So it changes the

baseline pixel gradients Exactly If the model

learned the subtle signs of mortality by primarily analyzing Fitzpatrick skin types one through three it is blind It

is completely blind

to everyone else It might completely misinterpret

the signs on faces with darker skin tones The baseline for what

a healthy nasal bridge looks like varies wildly across global populations Of course Different bone

structures cast

different shadows If the training data

is localized and biased the results are different

The resulting

diagnostic tool

is inherently biased And the stakes

here are quite literally life and death Yeah If the

model

hallucinates the risk factor on a patient of color because it misinterprets a natural shadow The clinical consequences

could

be severe Radiation oncologist Raymond Mack

was involved in the research And he explicitly

acknowledged this limitation He did He

stated clearly that broader more diverse data is needed And the team

is working to expand that training set right now They are Actual

clinical testing

is still entirely ahead of them We must also

consider the bioethics of this capability We really must

Matthew Liao was

a prominent bioethicist at New York University He was not

involved in this specific study No But he

raised a massive red flag regarding how this technology might be implemented in the real world He warned

that a mistaken estimate from this AI could lead to a patient being denied needed care Yes This

is the nightmare scenario for algorithmic medicine Imagine a

hospital setting An algorithm

mistakenly

flags a patient as having a very high 90 day mortality risk Right A

medical

board or an insurance underwriter looks at that score And what

do they

do They might decide that

aggressive curative treatment is no longer a viable financial or medical investment They might push

the patient toward palliative care Exactly And all

of

this could be based on a flawed facial scan The algorithmic score

becomes a self fulfilling prophecy It is

terrifying And Liao

also pointed out the risk of misuse outside the hospital walls Right What

happens

if insurers or employers get their hands on this kind of technology The corporate

applications are deeply concerning Imagine an

employer

requires a headshot for your company badge Which happens all

the time Right And

they

run it through a proprietary survival model Oh wow

Suddenly your

life insurance premiums go up Or you

are passed over for a promotion because an algorithm decided your nasal bridge indicates underlying frailty That is

the dystopian application It really

is It strips

all the context from the patient It reduces

a human being to a statistical risk vector based on

a photograph Which is

exactly why the Mass General Brigham researchers are heavily emphasizing their proposed aim They are

being very careful about this They are

not trying to build an autonomous diagnostic oracle No This

technology

is strictly meant to support clinicians The goal

is to make physical assessments more consistent Right Currently

two different doctors might look at the same patient and have very different subjective opinions on their overall frailty So

the AI just provides a standardized objective data point It

is meant to prompt a doctor to look closer It is

absolutely not meant to replace a physician's judgment Or make

the final call on the treatment plan It

is an assistive tool It

is an extra layer of screening So

what should you watch for next

You need to watch for the results of the expanded trials

We really need to see how this technology actually performs when it is exposed to diverse global populations

We also need to see the results of controlled clinical testing in real hospital environments

And the scientific community is still waiting to find out the biological reason why those specific facial regions correlate with mortality

Exactly

Until we understand the underlying physiological mechanism

it remains a fascinating piece of research

but not a diagnostic reality

That explicitly closes our lead story on medical AI

So in other

news We are seeing how AI can autonomously extract invisible biological signals from our

faces

Right But what happens when that same autonomous pattern recognition is pointed at our digital

infrastructure It gets ugly

fast It stops looking for

wrinkles It starts looking for

vulnerabilities We are looking at a massive cybersecurity

campaign

Reuters recently traced this operation to a Chinese speaking

hacker

And this attacker did not use traditional methods

They used autonomous AI agents to probe more than 1 200

websites Over 1

000 websites They successfully breached sites belonging to more than

100 companies

They did And this was not a highly manual keyboard

smashing operation Not at all The hacker used an open source AI problem solving platform

called Cairn The way this campaign was discovered is highly

unusual though It

really is The attacker made a massive operational

security error They

got sloppy

They did They actually left their command and control panels exposed to the

public internet Snorri Fagerlund is a principal threat researcher at a cybersecurity company

called

Norforge Right And he stumbled across these unsecured

Cairn interfaces Which is a huge stroke of luck

for defenders That discovery allowed him to essentially look over the

hacker's shoulder He watched the AI agents work through the intrusions in

real time The campaign started around

September 25th Yes And it heavily targeted

e commerce sites The victims were located in

the

United States Canada

and Canada Australia and Japan

So

the

scale

is massive but the input

was incredibly minimal Here is the part that completely changes

the threat landscape

It really does The hacker only wrote about 100 Chinese

characters of

instructions One hundred characters That is less text than a

standard social media post Those initial instructions

were incredibly high level

What did they say They simply set the preferred intrusion tactics

and the

target parameters And then From there the humans stepped back The

agents took over completely They

worked autonomously for hours with

almost no human direction Imagine you are running

an

e commerce site Normally defending your network is about

stopping

a human Right A human burglar has to walk up to

your digital house They

check the windows They try to

pick the lock And

if it fails

They get frustrated

They get tired And they move to

the next house It

is an

exhausting linear process But this AI

is entirely

different Entirely different This is a burglar who can simultaneously jiggle the doorknobs of 1 200 houses

for

24 hours straight They never need

a coffee break They

never lose focus And they never suffer

from alert fatigue And they

are

highly organized

too They are The Karen platform provides an abstraction layer

for the AI It sets up the work around a starting point and

a defined goal The agents actually use a shared project board

within the platform That shared

memory is

critical Very critical They record

what they discover They post their intended next steps

to the board So if one agent finds a specific version number for

a

server or software Another agent can read that clue and use it to search

for

known exploits Exactly They are not working in isolation They are actively collaborating to break

into your network There is a terrifying example of this from

the Reuters report

There really is In one specific case the agents were attacking a

company's public facing application They

hammered the front door And they

found absolutely no weakness

The application was secure A simple automated script would fail

at this point It would

just error out A traditional vulnerability scanner would log the

failure and stop But these agents

are

not scripts No They possess

lateral thinking

capabilities They

pivoted They did They started searching digital records

for related infrastructure They

looked for subdomains They looked for

forgotten developer

portals And eventually they found an exposed control panel connected to

the target company But finding the panel was only

the

first step They then spent hours using trial and error against that

specific interface They read the

error messages They adjusted their payloads based on the

server's responses And they eventually discovered a flaw in how that panel

processed data They manipulated that exact flaw to make the server execute

malicious commands They achieved remote

code execution They hit a

solid wall

stepped back found a side window and kept pushing until the

glass broke The persistence is

just incredible And apparently this persistence is

incredibly cheap How cheap are

we talking Well Fagerlund analyzed the token throughput on the

exposed

panels Okay He estimated that a successful intrusion took about 23 hours of continuous agent

work And the

cost He estimated the computational cost in AI tokens for that success was roughly 50

50 To

breach a major corporate network That

completely upends the economics of cybersecurity It

really does If

an attacker only needs 50 to fund an advanced persistent threat the

barrier to entry drops to zero Across

the entire campaign the

estimated token cost was about 5 000

For 5

000 the attacker compromised over 100

companies

However we must be very careful with those financial

numbers We

must NorForge explicitly stated that they lacked actual usage

data They do not have the hacker's credit card

receipts Or API billing

statements

No These dollar figures are estimates based entirely on pricing

assumptions The researchers noted that a worker name in the system suggested the campaign used Claude code as the agent

framework And they believe it was running on top of a

DeepSeek language model DeepSeek is

known in the industry for being highly capable and extremely inexpensive But these

dollar figures are just inferences of AI consumption Based on cash

tokens and assumed runtimes They're not a verified

budget We do not know exactly

what the attacker paid But we do know who

got hit We do Reuters identified some

major corporate

names caught in the dragnet Deere Shark Ninja And Cummins

were among the affected companies But the actual

damage remains very murky Deere stated they

took immediate remediation actions They claim they

had no indication that sensitive customer or corporate information was accessed Shark Ninja said

a standalone product support site was hit But the threat

was contained before it reached core systems Cummins is currently

investigating a potential attack on a website operated by a third party vendor Which perfectly highlights

the supply chain risk Right The AI

does

not have to breach the main corporate fortress It just has to

find the weakest vendor connected to the network That is exactly what

you need to watch next Visibility is the massive

unanswered question The attacker's control panels

were wide

open for researchers to see Yet the victim companies

are still struggling to determine exactly what data if any was extracted

from their

systems Even when defenders can

watch the agents at work the scale is overwhelming

Agent workflows can escalate

an intrusion so fast that tracing the extracted data through the server logs becomes incredibly difficult The automation just outpaces

the forensic analysis

Meanwhile AI autonomy is

not

just being used to attack us It is also being

deployed defensively We are seeing autonomous

agents designed specifically to waste the time of human scammers This brings us to

a fascinating Australian startup called Apate ai They have built

an army of AI decoys These agents are

programmed to pose as vulnerable gullible victims And

the company

has scaled this operation massively They report that

their bots have engaged human scammers in two and a half million distinct conversations Two and a

half million If you string

all that active conversational time together it amounts to

17 years of cumulative engagement 17 continuous years

of criminals talking to robots That is incredible

The origin story

of Apate is fantastic too Oh yeah Tell

it The

founder is

Dali Kafar He was at

a family picnic when he received a typical scam phone call We all get

those Right But instead

of

hanging up he decided to

just keep the caller on the line He played along

He feigned confusion

He kept the

scammer talking for 44 minutes It ruined his

afternoon picnic It consumed his

personal time But it sparked

a brilliant idea He realized

he could automate that exact distraction So he

started researching the concept at Sydney's Macquarie University And Apate

was officially founded in 2024 to commercialize the research Now they deploy thousands

of

concurrent agents These are not simple

interactive voice response menus either No These agents use

complex system prompts to generate varied personas They use different accents

They speak multiple languages

They answer suspicious inbound

calls They reply to phishing

texts They click on malicious

links They even proactively search

online forums and fake investment groups under fabricated identities They actively bait the

hook

If we connect this to

the bigger picture the goal is not just

to annoy criminals Wasting their time is a

nice side effect But the true objective is

shifting the power dynamic in cybersecurity Right These agents are given specific

proactive missions by their corporate

clients Clients like TPG Telecom and

the Commonwealth Bank of Australia are using this service TPG alone has diverted nearly 590 000

suspicious calls directly to these AI agents The agents are designed to extract actionable fraud intelligence They pose