# Anthropic: Enterprise Growth, Mythos Risks, and Pentagon Friction

**Podcast:** FT Tech Tonic
**Published:** 2026-05-06

## Transcript

Anthropic does a lot of safety research that might seem quite unusual or strange, especially to somebody who's not in the AI world.
This is Christina Criddle, who writes about artificial intelligence for the FT in San Francisco.
She's talking to me about Anthropic, one of the world's leading AI labs.
Their model welfare team basically explores the welfare of the models.
It's trying to work out whether AI systems experience behaviours that are like anxiety, distress, panic, especially when they are being put to different tasks.
Anthropic likes to portray itself as the good guy of AI.
The company worries about the impact that AI could have on the world, the damage it might do, and even, it seems, the feelings of the AI models themselves.
There was actually a paper route.
showing that tasks like trying to jailbreak or create a cyber attack made the model unhappy, whereas like creative writing would make a model happy.
And the idea behind this is like, obviously it's hypothetical, but if there's a slight chance that these models can grow to resent us because we're putting them to task all the time, we're making them work around the clock, if they grow to resent us and then they have super intelligence and all this power and control and agency.
over our financial institutions, our credit card details, don't we want them to maybe like us?
This is Tectonic from the Financial Times.
I'm Murad Ahmed, the FT's technology news editor.
A handful of Silicon Valley companies are vying to lead the world in artificial intelligence.
OpenAI, Anthropic, Meta, XAI, and Google DeepMind.
All of them are hoping to win the AI race and take control of our AI future.
In this season, I'll talk to the FT's expert tech reporters about each of these companies, and I want to start with Anthropic.
It's the AI lab set up by Dario Amadei and other former OpenAI researchers.
It's cast itself as OpenAI's more ethical alternative.
Amadei frequently warns about the dangers of super-intelligent AI.
But at the same time, Anthropic is rolling out increasingly powerful models.
The latest is called Mythos.
claimed to be so good at hacking it couldn't be released to the public.
So can Anthropic beat its arch-rival OpenAI, win the battle of the AI labs, and still claim to be the good guys of AI?
To understand where Anthropic stands in the AI race, I spoke to Christina Criddle, who you just heard from, and to John Thornhill, the FT's innovation editor.
John recently sat down for a lunch with the FT interview with Dario Amadei.
I was very keen to talk to Anthropic because they very much are the kind of company of the moment.
They've released these incredibly popular coding tools.
They've released this incredibly powerful model called Mythos, which has shaken up the kind of cybersecurity world.
So I was very interested to see how the company is run.
Dario Amadei, he came across as an incredibly focused, intense individual who is absolutely kind of laser focused on trying to develop super intelligence.
He's very much the kind of front runner, I think, in the field of AI at the moment.
And it's also an extraordinary company in terms of the kind of revenue it's generating.
It's one of the fastest growing companies in history.
So he's definitely someone worth interviewing.
So lots to talk about with Anthropic, but let's start with this latest model, Mythos.
It's caused a big stir.
Christina, what's the story here?
So Mythos is Anthropics.
It was meant to be a general purpose model.
So it was meant to be the next version of Claude that everyone would get their hands on.
But then in testing, Anthropic realized that it was very capable at cybersecurity.
And there were some of these capabilities which Anthropic thought were too risky to release to the general public or release very widely.
And that's because the software is very good at detecting vulnerabilities.
So trying to find bugs or flaws in digital infrastructure.
And then the second part that it's really good at is generating exploits.
So kind of workarounds for hackers or cyber attackers to take advantage of those vulnerabilities in software.
I kind of think that the first part...
already existed in a lot of models and especially even previous Anthropic models.
But it's that second part, the exploits, that is really significant.
And so Anthropic decided to release it to just like a select amount of vetted partners, basically only put this in the hands of companies that it could trust, giving it the chance to detect the vulnerabilities and try and patch them before any bad actors could take advantage of them.
And it's fair to say there's been quite the freak out over Mythos.
We've had heads of central banks, government ministers, regulators, companies all saying a variation of they are scared of it.
They want to be able to access it to see if they can test it as well.
How warranted is that reaction, do you think, given what we know about Mythos at the moment?
I mean, this was absolutely a moment where it sort of flashed into the consciousness of regulators, governments, banks, different organizations who saw how powerful AI can be for cybersecurity.
I think we already knew, at least in the AI world, that AI could be used as a very powerful tool for identifying software flaws.
But this seems to have put it on the map.
And actually, Anthropic has been accused of stirring up fear and capabilities to try and boost hype for this product, make it seem like other models aren't capable of it, maybe because they want to sell their product more.
There are other models which are very capable as well.
And even Anthropic's previous model, which is generally available, can do a lot of this stuff too.
So cynics will say this is maybe a marketing ploy to hype up their technology.
get more people to pay for their products.
Christine, I'm glad you brought up the issue of hype because every single time we've written about Mythos over the last week or so, there have been plenty of commenters saying that the FT's operating as part of the marketing arm for Anthropic.
I think...
Actually, so many people are wondering whether or not it has this effect on our cyber defences.
And if it does, then people need to prepare.
And that's why there's been so much conversation about it.
I mean, John, in your columns, you've often reflected on the power of these models.
What do you think about the Mythos release and what it means for all of us?
Well, I did put to Amadei the point that a lot of people are saying that this is marketing, to which he laughed.
And it's rather a peculiar kind of marketing tool as well to kind of try to stir up fears to show how powerful your models are, I think.
But I think more generally, there is a lot of debate about whether it's right that a private company should have such capabilities.
And there's been a bit of a debate about are they usurping the power of nation states and should companies like Anthropik even be nationalized because they challenge the security of a state.
And his response to that was that it would be too dangerous for any one entity, whether it was a private sector company or a government, to control this in isolation.
And therefore, that's why they set up this project, where they are trying to share a lot of this technology and try to fix a lot of the flaws that they see in the other systems before it becomes a kind of systemic problem.
And he hopes that that could be a model for how governance could evolve in future.
We are developing these fantastically powerful models now.
Their regulation is lagging a long way behind.
And so are there new forms of governance that the industry in collaboration with governments can devise that would help supervise and scrutinize these models before they do terrible damage?
I think it's a really nice idea, but it's maybe after the facts.
Like this technology was already being used for cyber attacks.
It has definitely ramped up with Mythos, but it already was out there.
And Anthropic contributed to that with its previous model, which it did test for these capabilities too.
Anthropic's done a lot of research where suspected nation state actors have used clawed code to try and make attempts on...
governments, healthcare institutions, financial institutions using its products.
And for me, what really is interesting is that Anthropic was founded with the mission of being very safe.
And it does do all of this safety work and reporting.
But Mythos just shows that they're willing to put software out into the world that might not be entirely safe.
And yes, they've done that to a limited amount of partners.
it could really have serious repercussions for global security.
And we also reported that there was unauthorized access to this Mythos model as well, which was a Bloomberg investigation found that a group of people did manage to access Mythos outside of this program.
So there are serious concerns about whether Anthropic can even make sure that this technology does not get into the wrong hands.
Okay, so still questions about the hype around the Mythos model, but it has put Anthropic in the spotlight and it's actually only the latest in a series of impressive models that have come out of Anthropic.
Christina, give us a bit of background here on Anthropic and how it's measured up against its arch-rival OpenAI.
So Anthropic was founded by former OpenAI employees and they wanted to create a company that was more specifically focused on making sure that these models are safe.
And you can see that through a lot of the work that they do, a lot of the testing and research that they publish.
And their models are at the top of the leaderboards.
They're often seen as state of the art.
They're often seen as overtaking OpenAI now in most of the benchmarks.
And they've really had some step changes where they focused on enterprise from the beginning.
They created this clawed code product, which was incredibly good at coding and really shook.
software engineers into realizing that AI was going to completely transform their jobs.
That was one moment.
And then they created Claude Cowork, which was a less technical version of Claude Code and was meant for anyone to use in their work life.
And that seemed to impact the legal sector quite a lot.
And so it's had these product launches where it has really shaken markets.
It's made...
People around the world who maybe aren't using AI all the time realize how impactful this is going to be for their careers.
And so I think by focusing on enterprise and focusing on work applications, it's managed to have a series of hits.
And then in terms of its roots and safety, I think that's helped it position itself as a friendlier company than perhaps other rivals.
And it's very thoughtful in the way that it delivers its messaging.
It's tried to paint itself as the good guy.
Well, it's painted itself as a good guy against a particular opponent in OpenAI.
Like you said, Christina, it came out of OpenAI.
John, when you spoke to Amadei, did you get the sense of that animus and that competitive element with OpenAI still exists?
I know you tried to press him on this and I don't know how far you got.
Well, he was very well.
prepped, I feel, because he was very resistant to personalizing any of the kind of issues that we've just been talking about.
He did kind of obliquely refer to chaotically oriented actors or chaos agents, as he called them, and then suggested that the FT readers might guess who they are.
But he was pretty critical of some of the billionaires in particular, in terms of their willingness to push the technology beyond what he thought was safe.
And so again, to pick up on Christina's point, they're very much positioning themselves as the good guys, although they have attracted very fierce criticism from other quarters in Silicon Valley.
This desire to be the good guys.
There have been other companies that have come along in the past.
Google famously had its motto, don't be evil, and that's long gone.
They became a very conventional company.
Christina, what do you think the chances are that they'll be able to maintain this position as the good guys or as they need more and more money for all the computing power that you're talking about?
Surely they're going to have to do some evil things.
They're going to have to do quite a lot more that is driven by the profit incentive.
Evil is a strong word, but I think...
When you look at Anthropic, of all of the AI labs, it does seem to have these very ideological employees.
They're motivated by wanting to do good.
Often they do devote a portion of their salary to charities.
And in a world where people are getting paid ridiculous amounts of money to work at AI companies, I think that is significant.
Anthropic also has a very high retention rate.
People don't tend to leave.
And so there must be something in the messaging or the ethos or the culture there that is making people stay.
But I do think that there does seem to be this commercialization of Anthropic happening at the moment.
It definitely is ramping up in terms of its revenues, in terms of its user base and the products that it's releasing.
And so you have to ask whether...
Some of these more ideologically driven work, like their model welfare group, which looks at whether models are having good well-being, whether they're panicking or experiencing something like anxiety, whether those kind of organizations will still be prioritized, that kind of experimental research as well.
When you do have this fast pace of trying to create a product that keeps up with the high demand that it's experiencing, Can it do both at the same time?
We've seen OpenAI also struggle with this.
And so I think that's going to be a big problem for AI companies to retain staff, but also make money.
Okay, let's take a quick break.
And when we come back, we'll talk about this argument Anthropic has had with the Pentagon and what impact it might have on the company's ambitions.
But with AI, which the development is always faster, increase the requirements of security.
And this earlier than before.
If things like Security and Compliance are really important, it allows growth.
If you wait for a long time, it breaks it off.
That's why many Startups early on WANTA and grow up.
And if it's going to be an audit, don't go in months.
Now start on WANTA.com.
John, what happened with Anthropic and the Pentagon?
They signed a deal allowing the government to use their technology, then seemed to object to how it was being used.
Well, as you say, Anthropic has signed a deal with the Department of Defense, or now war, and their models, Lord, has been used by the Pentagon quite extensively.
And it was reported that it was quite instrumental in the capture of Maduro, Venezuela's leader, for example.
technology to the military.
However, the Pentagon were trying to revise the terms of their deals with a lot of the kind of tech companies and wanted to have this kind of broader phrase that the models should be used for all lawful purposes.
And Anthropic kicked off about that in particular because they thought that in a certain set of narrow cases, it was possible that the use of the technology could infringe.
democratic rights, and in particular, domestic mass surveillance, and also for the development of lethal autonomous weapon systems or killer robots.
So they got into a big face-off with the Pentagon about this.
The Pentagon said, frankly, it was ridiculous that a private sector company was trying to dictate to a democratic government how the technology should be used for lawful purposes, but Anthropic stuck to their guns on this.
And so the Pentagon And the Trump administration declared that Anthropic was a supply chain risk, which is a very serious accusation against the company.
And it would have then led to the stripping of all federal contracts with Anthropic.
Anthropic have appealed this, and it's now currently going through the courts.
But I think if they cannot overturn that...
I think it would be very serious for Anthropic.
It's very hard to see how they could get an IPO away, in my view, if they're actually in a live dispute with the government.
It is remarkable that Anthropic thought that they could dictate terms in that way to the Pentagon, not only in the use of technology, but in this day and age with the Trump administration being so on the front foot and aggressive.
The rest of Silicon...
Valley appears to have kowtowed to the Trump administration, or at least maintain ties and not anger it.
What do you think it all says about Anthropic's place in the Valley, almost, or the way the company thinks about the relationship with government?
I think there's clearly a split in Silicon Valley at the moment.
I mean, as Christina was saying earlier, Anthropic is infused with this kind of strong streak of idealism.
They are concerned that these tools can be misused and abused, and they really don't want to be part of that.
It hasn't stopped OpenAI and Google now, it seems, picking up some of the contracts with the Department of War.
But it has also led to quite a backlash from the employees of all of these companies who are saying that they are also worried about how these systems can be used.
And it should be said, though, The Pentagon has tried to portray Amadei as some kind of peacenik.
And in some respects, he's actually very hawkish.
He's very hawkish on China.
He really does think that human level intelligence ought to be developed by democratic societies and is very worried about what happens if China was to get there first.
And secondly, he's also quite kind of excited, as he put it, that the US and other democratic societies could use...
Mythos to help dissolve, to use his phrase, authoritarian regimes.
He sees Mythos as quite a powerful weapon to help defend democracies such as Ukraine and Taiwan and counter potential adversaries such as Russia and China.
So it's quite a complicated story, both within Anthropik itself and more generally in Silicon Valley.
Christina, John mentioned Anthropik's plans for an IPO, going public, floating the company on the stock market.
Anthropic, like a lot of AI labs, doesn't make a profit yet.
But it's clear that it wants to become one of the biggest companies in the world, right?
It's a very competitive company.
And yes, it is a company and it's trying to make money.
And you can see that in the products it's creating in the success of Claude Code and Cowork.
And now with Mythos, these are very business aligned products.
These aren't for general consumers.
From very early on said, we're going to chase enterprise.
We're going to look at work applications.
And OpenAI is now also struggling to monetize some of its consumers.
release ads.
It's also had e-commerce launches, which have struggled to gain traction.
And now OpenAI has pivoted to enterprise as well.
But Anthropic has already been there for quite some time.
And so whether it's getting that return on investment first, because it's been there for longer, means that it might even win this race.
And I think what's extraordinary about Anthropic is the rate of growth of the revenue generation in 23.
They were making about $100 million of revenue.
In 2024, they got to about $1 billion of revenue.
And last year, they were close to $10 billion of revenue.
And given the run rate of recent months, it's not inconceivable that they could even reach $100 billion this year and maintain that kind of 10x annual increase, which would be an astonishing rate of growth and would be obviously a very good backdrop for any IPO that they were trying to stage.
And they also have a kind of thousand customers who are paying them a million dollars a year for their services.
The big question I think all investors will then focus on is what is the rate of growth going forward?
Trees don't grow to the sky.
Surely the rate of growth will tail off and other people will then start eroding their advantage in the enterprise space.
But that will be the critical question, I think, at the time of any kind of public flotation.
If it does float, will Anthropic just become another big tech giant?
And is that inevitable if you want to lead the AI revolution?
Yeah, I think there is this incredible tension at the heart of Anthropic.
And it's, as Christina was saying, it's the heart of some of the other companies as well, in particular, kind of DeepMind and Google, that if you want to remain at the forefront of the technology, you have to have...
massive amounts of capital, massive amounts of compute power, and massive amounts of data.
And the only really way that you're going to do that is to have a very convincing financial story or be backed by a very large company, as in the case of DeepMind and Google.
And so there is this kind of inexorable logic to the way that the company is developing.
They have started out, I think, with incredibly noble purpose.
The more money they take on, the more shareholders they acquire, the more responsibilities they have to return that money to their investors, the more they are going to become more like a normal company in the way that you suggest.
So I think that's the real tension now.
Can they combine both being a very profitable company that justifies a very large IPO and generate that money in the future to return to investors at the same time as...
being aligned with the safety concerns that they espouse.
What could stop Anthropic?
What are the potential stumbling blocks?
I think also the models are developing so fast that you just cannot see what's around the corner.
For the moment, Anthropic has got a lot of momentum, but in six months' time, one of their competitors might come out with a model that kind of blows away their capabilities.
And in particular...
Some of the big tech companies in particular, Google, Amazon, and Meta, have such torrential cash flows from their other operations that they can keep.
investing in developing these models for open and AI and Anthropic, which are startups, which are still dependent on private funding.
It's a very different game.
And so it's a question of whether any of these companies have a moat around their business and are able to sustain the advantage they have at the moment, or whether that just gets competed away.
Christina, last word to you.
What do you think Anthropic is going to be like as a company?
And how likely do you think it's going to be?
the winner or a winner in the AI race?
Well, if you take these AI companies at their word, businesses are going to look very different in the future.
People here talk about having one person unicorns, where it's just one person and their AI systems helping them raise all of this money and create all of this revenue.
So maybe these companies will look a lot slimmer, a lot leaner, and they will just be filled with AI agents who are creating all of the gains for them.
I do think because Anthropic, OpenAI, Google DeepMind, these companies have been at the forefront of creating the models and the technology.
It's very hard for other actors to catch up and create these very large language models, which need so much compute to get off the ground and then to sustain.
And with the IPO, OpenAI has also said it wants to IPO, as has Elon Musk's SpaceX, which now has XAI.
the AI lab as part of it.
And so I think there is this competitive race.
All of them are racing to try and get to IPO first.
Who will get there?
I don't know.
And so I think behind all of this really important work, pushing the capabilities of AI, you do fundamentally have these men who are leading these AI companies who are highly competitive, who all know each other as well.
And they want to win the race.
And they truly believe that they are the right stewards for this technology to make sure it plays out in our lives, in our businesses, as best as possible.
That was Christina Criddle in San Francisco and John Thornhill, the FT's Innovation Editor, talking to me about Anthropic, thanks to both of them.
Startups like Anthropic and its bitter rival OpenAI have made the running so far in the AI race.
But next week, we'll be talking about the big tech giant looming up behind them.
Google has loads of cash, masses of data, and its AI lab DeepMind has some of the best AI brains in the business.
So if we're asking who will be the Google of AI, is the answer, well, Google?
That's next week on Tectonic.
But before you go, we'd like your help.
We want Tectonic listeners like you to give us some feedback, so we've put together a short survey.
It's your chance to tell us what you like, what you don't like, what you'd like us to do more of.
There's a link in the show notes, and if you fill it in, you'll help us make a better podcast, and you'll also get the chance to win a pair of Bose wireless headphones to make listening to Tectonic even better.
So go to the show notes for the link to the survey and the T's and C's.
Tectonic was hosted by me, Murad Ahmed.
The producer was Edwin Lane.
Production assistance by Michaela Tindira.
When your options are limited, so are your opportunities.
At SIBO, the global exchange that pioneered options trading, we offer more ways to move with the market.
From VIX and SPX options to global market data solutions, SIBO helps investors diversify, manage risk and stay ahead of whatever the market does next.
SIBO.
Life is better with options.
Your investments could be too.
There are risks associated with SIBO Company products.
Review the disclosures and disclaimers at SIBO.com slash US underscore disclaimers.
The latest episode of the Next 5 podcast is all about fraud versus reality in the age of AI.
Simon Miller at CIFAS joins me.
AI has made scanning as scanners emotionally intelligent.
As does Gareth Murray at Monzo.
The constant arms race between the bank and the fraudster.
And Uber Behagel at Verif.
When one data leak happened on the internet, how many times is it sold on the dark web?
Listen to the full episode of The Next 5 wherever you get your podcasts.
Enjoy.