In partnership with
Become An AI Expert In Just 5 Minutes
If you’re a decision maker at your company, you need to be on the bleeding edge of, well, everything. But before you go signing up for seminars, conferences, lunch ‘n learns, and all that jazz, just know there’s a far better (and simpler) way: Subscribing to The Deep View.
This daily newsletter condenses everything you need to know about the latest and greatest AI developments into a 5-minute read. Squeeze it into your morning coffee break and before you know it, you’ll be an expert too.
Subscribe right here. It’s totally free, wildly informative, and trusted by 600,000+ readers at Google, Meta, Microsoft, and beyond.
A few days ago, a story spread quickly across X.
The headline was dramatic.
"Anthropic's Mythos AI breached classified systems in hours during an NSA test."
At first glance, it sounded terrifying.
Did an AI secretly hack into government networks?
Was national security compromised?
Not exactly.
The truth is less scary, but much more interesting.
And it tells us a lot about where AI is heading.
The story comes from comments made by U.S. Senator Mark Warner during a discussion on AI and national security.
According to reports, Warner shared remarks from General Joshua Rudd, who leads the NSA and U.S. Cyber Command.
Rudd described a red-teaming exercise involving Mythos, an unreleased model from Anthropic.
The goal was simple.
Test how well existing security systems could defend themselves against a very capable AI.
The environment was controlled.
The AI was allowed to use certain tools.
Researchers observed everything.
And then something unexpected happened.
The model found ways into highly protected systems much faster than people expected.
Not in weeks.
Not in days.
In just a few hours.
Before going further, let's understand one important thing.
This was not a real cyberattack.
No foreign government was hacked.
No military secrets were stolen.
No classified network was secretly compromised.
This was a simulation.
Think of it like hiring expert thieves to test the security of a bank.
You give them permission.
You watch what they do.
And you learn where your defenses are weak.
That is exactly what red teaming is.
Companies do it.
Governments do it.
And now, increasingly, they are using AI to do it.
Still, the results shocked many people.
Because AI models are getting better at things that were once considered uniquely human.
Reasoning.
Planning.
Problem solving.
Searching for weaknesses.
Combining many small clues into a larger strategy.
The scary part is not that AI can write essays anymore.
The scary part is that AI can act.
It can try things.
Fail.
Learn.
Try again.
And sometimes it can do this faster than humans.
Shortly after the story spread, journalist Shashank Joshi clarified an important detail.
The model's success depended on specific tools and setups.
In other words, Mythos was not magically breaking into any system on Earth.
It was operating in a controlled environment with access to tools designed for testing.
That context matters.
A lot.
Because headlines often make AI sound like science fiction.
Reality is usually more complicated.
But even after the clarification, one fact remained.
The AI was extremely capable.
Capable enough that senior officials felt the public should know about it.
This story arrives at a strange moment.
Governments around the world are becoming more powerful because of AI.
But they are also becoming more nervous.
The United States has been discussing restrictions on who can access advanced AI models.
Some officials worry that powerful systems could be used by hostile countries.
Others worry about cybercrime.
Some fear misinformation.
And some simply fear moving too slowly.
Because if AI can discover vulnerabilities in hours, what happens when these systems become even better?
What happens when millions of people have access to them?
What happens when open-source models catch up?
These are no longer science-fiction questions.
They are policy questions.
And they are being debated right now.
I find this story fascinating for another reason.
It shows the double nature of AI.
The same technology that can attack systems can also protect them.
Imagine an AI security guard.
It scans your network every second.
It notices unusual behavior.
It finds vulnerabilities before hackers do.
It patches software automatically.
It explains the risks in plain English.
That future sounds amazing.
Now imagine the opposite.
An AI attacker.
It never sleeps.
It tests millions of possibilities.
It coordinates attacks.
It writes malware.
It adapts faster than humans can react.
That future sounds terrifying.
The reality is that both futures may arrive together.
This is not the first time technology has forced us to rethink security.
When the internet became mainstream, people worried.
When smartphones arrived, people worried.
When cloud computing became popular, people worried.
Each time, society adapted.
New laws appeared.
New industries emerged.
New jobs were created.
Security improved.
The same will probably happen with AI.
But the speed feels different this time.
AI is improving every few months.
Capabilities that seemed impossible last year are becoming normal.
That is why governments, researchers, and companies are rushing to understand what these systems can do.
There is another lesson here.
The best way to prepare for powerful AI is not to ignore it.
It is to test it.
Aggressively.
Openly.
Responsibly.
That is exactly why red-teaming exists.
You want to know what can go wrong before real attackers discover it.
You want your own AI to attack your systems before someone else's AI does.
In cybersecurity, pretending a threat does not exist has never been a good strategy.
As readers, we should also be careful with headlines.
The phrase "AI breached classified systems" creates fear.
But the more important story is this:
AI is becoming good enough that governments are using it to test their most secure systems.
That is a remarkable milestone.
It means AI is moving beyond chatbots and image generators.
It is entering the world of defense.
National security.
Critical infrastructure.
And decisions that affect millions of people.
I don't think the lesson is to panic.
I think the lesson is to pay attention.
Because the future of AI may not be decided by who builds the smartest chatbot.
It may be decided by who learns to use AI safely.
Who builds better safeguards.
Who creates systems that humans can trust.
The Mythos experiment was not a warning that AI has taken over.
It was a glimpse into a future that is arriving faster than many expected.
And if an AI can expose weaknesses in hours today,
the question we should ask is not,
"Can AI break things?"
It clearly can.
The more important question is,
"Can we build the safeguards fast enough?"
Because the answer to that question may shape the next decade of technology.
—Sushila