Anthropic's Claude Mythos and the Future Trajectory of AI in Business

Citation: Katina Michael with Melinda James, "Anthropic's Claude Mythos and the Future Trajectory of AI in Business," ABC Illawarra Radio: Mornings, 15 April 2026, 8.45am, https://www.abc.net.au/listen/programs/illawarra-mornings/mornings/106555322

Host (Mel):
All right. You've probably heard of ChatGPT. Maybe you've heard of Claude. Do you use Grok? All these different forms of artificial intelligence. Well, have you heard of Mythos?

This is the new artificial intelligence model that has been announced by Anthropic.

But it's a bit scary for a couple of different reasons. One is that it has seemingly become so powerful that it has been able to do things and escape and do things that its creators did not expect it to be able to do.

The other thing that's arisen out of this is that Mythos has uncovered all sorts of software vulnerabilities—flaws in firewalls, etc.—in all the major operating systems and web browsers that we use all the time.

We don't know exactly where those security risks are because Anthropic isn't releasing them to the public, but probably for good reason, so that they can't be exploited by people with malicious intent.

It's all a bit scary, isn't it?

We're going to talk to Katina Michael in just a moment, our expert to help us make sense of all this.

But just quickly, Chris in Lake Heights. Hi, Mel.

Caller (Chris):
I've just started a small consulting business and I'm using AI quite a lot. I've managed to set up Claude, which is the Anthropic product, to develop a series of agents to function like a $70,000-a-year administrative assistant.

It manages my email, marketing, finance and tasks to prepare me for meetings. The power of this technology is incredible, but also very scary.

Host (Mel):
Anyone else use AI regularly? We'd love to hear from you. Or have you got any questions you don't quite understand about what it is?

Katina Michael is a professor at the University of Sydney Business School, the Director of the Technology and Digital program, and joins us now.

Katina, hi.

Katina Michael:
How are you, Melinda?

Host (Mel):
Oh, look, a little bit edgy after all this.

Look, I know I don't know a lot about this stuff. It confuses the heck out of me, but I'm deeply, deeply fascinated by it all.

I think people just need to conceptually think differently about AI in a way that they're maybe not used to thinking about ordinary computer programs.

But tell us about Mythos and Anthropic's announcement. Why does this have the world talking?

Katina Michael:
Well, this seems to be an escalation in the progression of AI.

When you look at it against other software engineering benchmarks and benchmark performance, these are just standard benchmarks in the industry. Mythos comes up incrementally better—not just by five or ten points, but more than that—against its closest competitor, OpenAI's ChatGPT, and even its own Claude latest version, Opus.

Mythos has purportedly been trained on 10 trillion parameters, which is extensive, and its reasoning ability, coding ability and cybersecurity capabilities are far better than any others.

In fact, when Mythos was pointed to about a thousand open-source domains and real-world applications, it came up with thousands of unknown vulnerabilities.

These are zero-day attacks. So, if they were to be exploited in any way, they could bring down some of the biggest web browsers, operating systems, cloud-provider software and other things.

It's done things autonomously that weren't expected to be done.

Previously, perhaps with Claude Opus, you might have required some coding ability and some knowledge of cybersecurity to create ransomware. This can do it automatically.

So, it detects vulnerabilities that have gone undetected for some 30 years in some cases, like OpenBSD's operating system.

It's even attacked Linux vulnerabilities, which we all thought was the safest operating system in the world, and it's chained them together to exploit and get to the root drive of passwords.

It's done all sorts of things with codecs that people have been studying for a decade or more.

So, it's demonstrated that it can outstrip human capacity.

Host (Mel):
Okay. And so, look, what we seem to have on our hands here, then, is this company Anthropic that has announced that it has developed this AI tool called Mythos that can do all this stuff, that has exposed all of those computer systems that the world relies upon—all our financial systems, everything, basically—and has exposed the fact that if this information got into the hands of the wrong people, they could hack the heck out of everything.

And so, where do we sit now? This is all being held in the hands of this one company. What's happening?

Katina Michael:
Well, some have called it an existential dread, but others have said it's a great marketing ploy by Anthropic—a sort of stunt, I wonder.

Host (Mel):
Possibly.

Katina Michael:
Possibly, but it's warranted.

We do have 40 companies that have already signed up to something called Project Glasswing, which offers them millions of dollars' worth of credits.

That's Microsoft, Apple, Google, Amazon, Cisco, Palo Alto Networks, NVIDIA, CrowdStrike and so many others, including foundations like the Linux Foundation and Apache Foundation.

But then we've also got the U.S. Treasury Secretary, Scott Bessent, calling in the banks and saying, "It's time to have crisis talks."

The Bank of England have also done that with insurance companies in the UK.

So, it's obvious that it's important, and there's a window of opportunity where we can say, "Let's exploit these capabilities. Let's use it from a defensive point of view. Let's block all those things, patch up where we have those vulnerabilities, and do something about it."

Host (Mel):
But in fact, what next? Will it be released to the public? Surely it will.

Well, what is happening? What's Anthropic said about whether they'll release it to, for example, financial institutions to let them know, "Here are your vulnerabilities. Here are where your clients' information—and let's face it, their money—is at risk of being stolen"?

What kind of obligations do they have? Can anyone force them to divulge this stuff? Where are we at?

It's all moving too quickly again, isn't it, Katina?

Katina Michael:
Yes. Well, we can quickly talk about DeepSeek, which we know has been used to create ransomware.

We know even Claude, in the Opus version, was used to create ransomware. About 30 organisations were hacked, but by people who knew how to code and who knew about cybersecurity.

Now this doesn't require that.

But really, it's the fact that you can't release it. That's a great sort of monetiser in itself. So, here we give it to you, not other competing companies that might attack you, or competing hacking entities.

But what you're doing is monetising that value as long as you can.

It's a change in the way we see things in this space, and I do think Anthropic and Dario Amodei have basically come out and gone, "Okay, now, OpenAI, you try and beat that."

But they will have to inevitably, as a trajectory, looking at how we will use AI in the future, release it to the public, I feel.

Host (Mel):
Yeah. But who and how is that all going to happen?

We know Anthropic—we won't get into that can of worms because it's a bit complicated—but they've got a damaged relationship with the current Trump administration as well.

So, it's all, yeah, a little bit scary.

But thank you so much for helping us try to make sense of it.

Katina Michael:
Thank you.

Host (Mel):
That's Katina Michael from the University of Sydney's Business School.

It's—it is a bit terrifying, isn't it?

Joseph's on the line. Hey, Joseph. How are you going?

Caller (Joseph):
Yeah, good mate. How are you doing?

Host (Mel):
Yeah, good. So, you work in this kind of area, do you?

Caller (Joseph):
Well, they pay me. I'm not sure "work" is the right word, but I'm definitely paid to be in this kind of area.

Host (Mel):
All right. So, what do you—I mean, is Mythos all the talk?

Caller (Joseph):
Well, as I said, it's in the name. It's a myth.

I mean, I'd like to show you my powers of flight. I'd like to show you the alien spacecraft, but it'll be really upsetting for you.

So, you know, they made a large number of claims.

Host (Mel):
And you can't make all this public, though, Joseph. You can't just put out there, "Here are all the vulnerabilities we found in firewalls and security patches and all the rest of it."

We found all these things that will enable hackers to exploit them.

Caller (Joseph):
Yeah, actually, this does look scary. But if I'm telling you I'm a scary person, you know...

And again, it falls in immediately after their fallout with the Trump administration over DoD use.

So now they say, "Well, hey, you know, aren't you upset that you didn't—I'm a lot more attractive boyfriend than you first thought I was. I've now got a faster car and I've got a lot more cash. Aren't you disappointed you didn't spend more time dating me?"

It's that jilted-boyfriend sort of response. It's that sort of, "Look at all these things we've done."

But we've not seen—and a lot of this stuff is all around claims.

Host (Mel):
So, how are we going to deal with these things, Joseph?

We've only got a couple of minutes before the news, but I feel deeply uneasy about corporations being in charge of this.

But then, given the state of the world, in some instances, I feel deeply uneasy about governments being in control of it.

Caller (Joseph):
Well, some governments we could probably give it to the Danish people to look after.

Host (Mel):
Oh, they do an all right job, wouldn't they?

Caller (Joseph):
Okay job, but no, no. It's probably better if some private organisations—and, yeah, then let them face the market for their bad behaviour.

But, yeah, look, it is scary. Happy to talk more about it when this comes up again.

But, you know, it can be managed. I mean, we've done scarier things in the past. Easy enough.

This is the nuclear—this is the Manhattan Project of our age.

We didn't blow ourselves up so far in the last 70 or 80 years. Maybe we can avoid it again for another 70 or 80 years with AI.

That's my take on it.

Host (Mel):
Okay. Thanks, Joseph.

Caller (Joseph):
Cheers, buddy. Have a great day, Mel.

Host (Mel):
Cheers. See you.

Oh, yeah. Is it all a bit over your head? I mean, how are you feeling about it? Do you keep up with all this stuff?

It really is the biggest change we're going to go through, and I think it's kind of understated in a lot of ways.