kedsley

breaking down AI threats before they break the news

[kedsley @kedsley]

a finance worker at engineering giant Arup wired $25 million after a video call where the CFO and every colleague on screen were AI deepfakes and the same con had already taken over $120 million from Google and Facebook

15 transfers to 5 accounts and nobody noticed for about a week

a Lithuanian man fooled Google and Facebook with fake Quanta Computer invoices and got 5 years in prison

the FBI counted more than $55 billion lost to fake boss emails in 10 years

it always runs the same script, a senior person, a secret deal and no time to check

Ian Goodfellow's 2014 paper set a forger AI against a detective AI and that game grew into today's deepfakes

in a 2025 iProov study only about 1 person in 1,000 spotted every fake

5 minutes that could save your company millions, call back on a number you look up yourself and bookmark this

[Jamieson O'Reilly @theonejvo]

https://t.co/85Kn8cUJq4

[kedsley @kedsley]

full paper:

https://t.co/JKvyjEzYRh

[kedsley @kedsley]

Google, Amazon and Nvidia poured $235 million into Hugging Face at a $4.5 billion valuation while three researchers hid 36.9 MB of working malware inside a famous AI model and all 58 antivirus engines on VirusTotal missed it

the 2021 EvilModel paper hid the payload in the last bits of each weight so the numbers barely move

an AlexNet with 60 million weights has room for about 180 MB of hidden payload

PyTorch pickle files are worse because they can run code the second you load them

JFrog found about 100 malicious models on the hub and one opened a reverse shell

Protect AI scanned 4.5 million model versions and flagged over 350,000 problems

use safetensors, check the hash and scan every model before you load it

5 minutes that change how you download AI models, save it before your next one ↓

[Jamieson O'Reilly @theonejvo]

https://t.co/85Kn8cUJq4

[kedsley @kedsley]

Apple swore its iPhone photo scanning AI would wrongly flag less than 1 in a trillion accounts a year and a developer cracked it open in two weeks

a few hours later a researcher forced two unrelated images onto one fingerprint

Google needed 9 quintillion computations to break SHA-1, while Apple's AI hash fell to an open source gradient trick

the birthday paradox explains why a 96 bit AI hash gives you only 48 bits of real safety

the same trick walks a photo of a cat to the exact hash of a dog

Apple dropped the whole plan 15 months later, watch this before you trust any AI with your security

[Jamieson O'Reilly @theonejvo]

https://t.co/85Kn8cUJq4

[kedsley @kedsley]

Researchers from Google DeepMind told ChatGPT to repeat the word poem forever, spent about $200, and walked away with more than 10,000 chunks of its real training data, including actual people’s emails and phone numbers

A month later The New York Times sued OpenAI and Microsoft for billions over articles GPT-4 could recite almost word for word, and the one inequality the industry now leans on, differential privacy, is the same math the US Census used for the 2020 count, this video shows how it works and how to test your own model for leaks

5 minutes, watch it before you paste anything private into a chatbot

[Jamieson O'Reilly @theonejvo]

https://t.co/85Kn8cUJq4

[kedsley @kedsley]

Google researchers made a top AI swear a panda was a gibbon with 99% confidence just by adding noise no human eye can see, and the man behind it, Ian Goodfellow, invented GANs and went on to direct machine learning at Apple

Then McAfee hackers put 2 inches of black tape on a 35 mph sign and a Tesla read 85 and started speeding up, while Tencent pulled a Tesla Model S toward the opposite lane with three tiny stickers on the road, and with Tesla worth over $1 trillion and going cameras only, every camera AI and every chatbot reading your screenshots shares the same weak spot from one line of math

5 minutes, bookmark & watch today before you trust a camera with your life

[S 🫧 @He1s_Sammy]

https://t.co/VT6ye4CuxW