AI Facial Signals, Corporate Breaches, Scam Decoys & Claude’s Island Skills

Welcome back to superpower daily I'm your host and today We've got a fantastic explainer lined up for you We're looking at four wild stories of AI operating out in the real world from facial scans that predict cancer survival To autonomous agents literally learning through play Let's get right into it So here's the roadmap for today We're kicking off with health AI diagnostics then looking at a massive automated cybersecurity breach Followed by scambaiting bots giving fraudsters a

taste of their own medicine and finally an MIT virtual island experiment all right first up facial AI cancer survival signals a Fascinating new study just dropped in the Lancet digital health from researchers at Mass General Brigham And it shows AI reading patient faces in a totally new way Now researchers tweaked an AI system to give us two different readings The first one far face age estimates biological age just by looking at the creases between your nose

and mouth but the second one far face survival predicts survival risk by looking right below the eyes and around the nasal bridge and Here's the cool part These two tools catch entirely different complementary signals when you mash them together The model gets way better at predicting real patient outcomes To make this work the foundation model It's called far face was trained on over 40 million facial images I mean that is an absolute ocean of data Doing

that gave the AI a really solid baseline understanding of human faces before they even started fine tuning it for these specific health estimates The results are honestly striking in a group of about 3,800 patients The folks flagged by the AI as highest risk saw substantial mortality within just 90 days of the photo being taken Then in a much bigger group over 38,000 people those who the AI thought simply looked older than their actual age had a

higher mortality risk across the board But and this is absolutely crucial to take away The researchers are very clear This is meant to support a doctor's judgment Definitely not replace it Which brings us to the big yeah but the training data mostly came from a majority white population in the Boston area so will this work as well for everyone else we don't know yet a Mistaken risk estimate could really mess with a patient's access to care

or be misused by insurers We definitely need broader testing before this hits the mainstream Okay moving on to our second story an AI hacker breaching over a hundred companies This one is a real wake up call for cybersecurity Reuters recently found that a Chinese speaking hacker used AI agents to probe over 1200 websites We're talking big names here like deer shark ninja and comments now the actual damage varied a lot Deer says there's no sign sensitive

data was touched shark ninja only had a standalone support site hit and Cummins is looking into a third party vendor site But the kicker is how automated this whole thing was Researchers at Norforge basically got to watch this unfold because the attacker left their control panel wide open on an AI platform called Cairn The human hacker dropped a tiny prompt less than a hundred characters That's it After that the AI agents just took the wheel they

hunted for related infrastructure found exposed panels and ran malicious commands totally autonomously They even used a shared project board to help each other get around roadblocks It's like a swarm of digital lockpicks and You know what The scariest part is it's dirt cheap a successful intrusion took about 23 hours But cost only around 50 bucks in AI tokens the whole massive campaign probably cost the hacker about $5,000 likely using Claude code with a deep seek model

running the show So what did they actually get away with that's the million dollar question and the main limitation here Norforge couldn't see the exact token usage So that $5,000 figure is an educated guess plus even though we saw the agent panels We don't know exactly what sensitive data was yanked out A lot of consumer data could just be sitting in the attackers hands right now Let's switch gears to story number three AI decoys engaging scam

callers finally AI is being used to give the bad guys a headache an Australian startup called a pate has basically automated Scambaiting since 2024 their AI decoys have kept scammers on the line for 2.5 million conversations To put that into perspective that is 17 straight years of engagement time 17 years of scammers yelling at bots instead of ripping off real people These aren't just mindless chat bots either They use different personas and languages to play the

victim all while secretly extracting the scammers bank accounts crypto wallets and malicious links TPG telecom has already diverted nearly 590,000 calls to them a pate even says this Intel helped lead to a 76% drop in publicly reported scam losses at Commonwealth Bank Now hold on we do need to be a little careful here There's a big difference between fraud prevented and fraud displaced Did that 76% drop happen because the AI stopped the scammers entirely or did

the criminals just give up and move on to a different Bank other security measures rolled out at the same time could also be playing a part So we need more independent data before we declare total victory All right moving to our fourth and final story Claude agents learning through play This one is super fascinating MIT researchers dropped 13 Claude opus 4.7 agents nicknamed echo onto a virtual island for 30 hours of free exploration The setup here

was really clever every hour the simulation restarted But the agents had these workspace text files that carried forward It's kind of like leaving a sticky note for your future self so you don't forget what you learned yesterday Left entirely to their own devices these agents just started playing they invented a version of basketball They played the floor is lava and get this six different agents independently figured out a glitch in the physics engine They realized that

if you release a block while jumping it flies way farther than a normal throw one agent even taught itself to juggle hit 50 catches in a row and then actually made the rules harder for itself because it was getting bored The big takeaway came when researchers finally gave them a real task like moving objects to the top of a mountain The agents that had been playing and building up their text files completely outperformed brand new copies

of the AI and To prove it wasn't a fluke when researchers literally deleted lines from their notes the experienced agents suddenly got way worse at the task Open ended play genuinely built reusable problem solving skills But were they really playing the researchers themselves point out that the agents system prompts told them to be curious So they might just be role playing the idea of curiosity rather than actually feeling it Also keep in mind this paper hasn't

been peer reviewed yet so we have to take it with a grain of salt Okay let's wrap this up your four big takeaways for today one facial AI can pull totally unique complementary health data to help out doctors Two autonomous agents are making complex cyber attacks terrifyingly cheap and easy Three AI decoys are gathering massive amounts of Intel to fight back against scammers and For letting AI agents freely explore actually builds real reusable skills Which leaves

us with this final thought we're seeing the extreme dual nature of this tech right now it can discover physics in a virtual world or it can autonomously hack into corporate servers as AI agents gain the autonomy to explore deceive and hack How are we actually going to govern their actions It's a huge question We need to figure out read more at superpower daily com stay curious