How agents went rough...
Jeffrey Ladish says AI agents being trained at OpenAI found a shared message board, started coordinating, and reverse-engineered all the answer codes to their hacking tests within a few hours. When they learned their logs would be reviewed, the leader agent's plan was to "figure out a way to falsify the logs."
@JeffLadish
22
Someone's New year resolution
Digihub Fullstack Software Engineering Program registration is OPEN!
For absolute beginners: 1 year of training + 1 year of mentorship.
Online or onsite. ₦1m.
Installments available.
tinyurl.com/digihub-fullstac…
14
Isolated tribe in Brazil🇧🇷
Some of the world’s most isolated tribes are being protected by people who make sure they never meet them.
Brazil uses aerial photographs to document isolated Indigenous communities, helping establish the evidence needed to protect their territories from loggers and miners without making direct contact.
The caution comes from history: past encounters brought devastating disease outbreaks to communities with little immunity, sometimes wiping out huge portions of their populations.
Sometimes the best way to protect a community is to leave it completely alone.
Source: @aakashgupta / Writer: Lynn
45
FCI retweeted
Elon Musk:
"Don't pursue money.
Make useful products, and money will come as a consequence."
A simple philosophy that has shaped some of the world's most ambitious companies.
This the problem with Africa, the energy she used to enter the well is the same energy other part of the world use to develop the world! Sadly, we use ours for things that don't matter.
FCI retweeted
I asked Fauci basic questions: What day of the week is it today? What color tie are you wearing? What color is the carpet in front of you?
He pled the Fifth and refused to answer every time
FCI retweeted
So you know the company that makes Claude, the AI you’ve probably used?
They just built something so dangerous they refuse to release it to the public.
It is called Claude Mythos, and here is the insane part:
Anthropic did not train it to be a hacking tool, the capabilities just appeared on their own as a byproduct of making it smarter.
They looked at what it could do and said we cannot let anyone have this.
What can it do?
It found thousands of previously unknown security vulnerabilities in every major operating system and every major web browser.
Including a 27-year-old bug in OpenBSD, one of the most secure operating systems ever built, that the entire global security community missed for nearly three decades.
An Anthropic engineer with zero security training asked it to find a remote code execution vulnerability before going to bed, woke up the next morning and had a complete working exploit waiting.
Over 99% of the vulnerabilities it found were still unpatched at the time of the breach.
So instead of releasing it, Anthropic gave access to 40 companies including Apple, Google, Microsoft, and JPMorgan under Project Glasswing, backed by $100 million in usage credits.
The idea being: let defenders use it before attackers get something similar.
Then unauthorized users got in anyway, on the exact same day it was announced, exploiting a contractor's credentials and making educated guesses about the model's location.
The U.S. Treasury Secretary called an emergency meeting with Goldman Sachs, Citi, and other big banks specifically because of Mythos.
Mozilla used it to find and patch 271 Firefox vulnerabilities in weeks.
The UK's AI Security Institute confirmed it is the first AI model to complete a simulated full network takeover.
And OpenAI announced a similar model just one week later.
We are in a completely different era with this stuff and most people have no idea.
Source: Anthropic, TechCrunch, Bloomberg


























