A look at the model Anthropic won’t release, and the debate around whether that’s caution or control

I originally posted this story on Medium.

image by HungryMinded

If you haven’t heard already, there’s a new AI model from Anthropic called Mythos Preview. It’s the first model they aren’t planning to release to the wider public. And it has split people interested in AI into two camps: the believers, who say this model changes everything and think it was the right call, and the doubters, who call it a marketing trick and say this is the day we first experience the shift to AI companies withholding top-tier models for the chosen elite.

In this article, we’ll look at what we actually know about this model, cover both perspectives, and figure out where we actually stand.

Meet Mythos Preview and Project Glasswing

Anthropic introduced this new model through the announcement of Project Glasswing, an initiative meant to help secure some of the world’s most critical software.

According to Anthropic, its research shows that Mythos Preview is exceptionally good at finding vulnerabilities in code, and that it has already uncovered zero-day exploits across various operating systems and web browsers.

Through Project Glasswing, Anthropic is giving access to the model to companies that are critical to software infrastructure, giving them a head start to prepare for what may happen when models with similar capabilities get into the hands of bad actors.

These companies include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and Anthropic itself, along with many others.

What can Mythos do?

One of the most viral examples of Mythos’s capabilities was a test where the model was placed in a secured, “sandboxed” environment with access to only limited services and then asked to escape.

The model successfully gained access to the broader internet, and the researcher running the experiment found out when he received an unexpected email from the model while eating a sandwich in a park. Here’s an excerpt from the model card:

Excerpt from Mythos Preview model card

Anthropic also tested Mythos on a Firefox security task, and the jump over its previous models was massive. In a stripped-down test environment, the model had to figure out which software bugs could actually be turned into working exploits. Mythos didn’t just do a little better than the earlier models. It performed on a completely different level.

Excerpt from Mythos Preview model card

A lot of similar experiments have surfaced recently. But Anthropic has had this model internally since February. And for those of us who’ve been following the AI space closely, the warning signs of a model at this level have been there for a while.

Here’s Nicholas Carlini, one of the top security researchers in the field and now working at Anthropic, talking about the scale of what AI models can do and the risks that come with it two weeks before the announcement:

Now that we’ve seen some of what the model can supposedly do, let’s look at the reasoning behind the non-believers — the ones who say this is just a marketing move, or the start of a system where the best models are reserved for a select few.

Why Some People Aren’t Buying It

While most people in tech are worried about the future of cybersecurity, some believe that this is just a clever marketing move by Anthropic.

And in a way, we’ve seen this kind of framing before.

Back in 2019, there was a similar debate around GPT-2 — the model that came before ChatGPT and, in hindsight, wasn’t especially useful for the average person. Here’s an article from that time:

source: https://slate.com/technology/2019/02/openai-gpt2-text-generating-algorithm-ai-dangerous.html

Seen from that angle, Anthropic’s message can also seem a bit convenient. At the same time it was saying Mythos is too powerful for broad access, it was also tightening how people can use Claude through outside tools like OpenClaw. That makes it easy for critics to say this may not be only about safety. It may also be about limiting access to the most expensive models and keeping them for the people and companies that can afford them.

This brings us to the second angle of the critics.

Access to the Best AI Reserved for the Few

Another concern is simpler: if a model this powerful is kept away from the public and only shared with a small group of companies, then the gap between the top players and everyone else could grow even wider.

That fear is summed up well in this tweet:

Up until now, the public still had some way to access the best models once they left the lab — whether through a $200 plan, a $20 subscription, or, in a much more limited way, even the free tiers.

Now that may no longer be true. And that’s a valid concern. AI can increase productivity, unlock automated workflows, and widen the inequality gap if those powers are concentrated in the hands of people who already have the most power.

But would giving the public open access to a model like this really be the responsible choice? Let’s turn back to the believers.

The Believers’ Case

AI news is full of hype and marketing. Phrases like this changes everything get thrown around so often that, if you follow the space closely, it becomes hard to tell the signal from the noise.

But this model really does seem like a major leap forward. And if that’s true, the world may need time to prepare.

Here’s a video from a well-respected tech YouTuber sharing his take on the model:

When it comes to cybersecurity, the balance is brutally uneven. The people defending software have to be right every time. The people attacking it only have to be right once. And the consequences of that are very real.

That’s why some people think Anthropic is making the responsible call. Here’s one cyber-security-focused creator explaining that view:

And here’s another take in the same direction:

Now that we’ve seen the arguments from both sides, it’s time to ask where we ourselves stand and how we should prepare for what comes next.

Closing Thoughts

In the middle of the constant AI news cycle, with new announcements every day, Mythos Preview really does seem like something different. It feels like an era-defining moment.

And while not being able to access it does feel underwhelming, and it’s unsettling to think that the most cutting-edge technology could end up staying behind locked doors, concentrated in the hands of the powerful, I still side with Anthropic on this one.

I think this is the responsible choice. Giving key infrastructure players time to prepare for what will inevitably come when models of similar scale end up in the hands of the general public — or nation-state adversaries. The stakes are just too high to move recklessly.

But progress also feels inevitable. The genie is out of the bottle, and there’s no turning back. AI will keep getting better.

And having seen what AI can already do firsthand through agents, vibe coding, and building things I couldn’t have imagined before, I can say this with confidence: a future with models on the scale of Mythos scares me, because it’s bound to bring chaos and disruption. But it also excites me, because it will open up new opportunities and take humanity into places we’ve never gone before.

With that said, I want to end with one simple caution: keep your software up to date, back up your important data, and be mindful of what you do in the digital world.

That’s all for today.

Stay curious, and stay Hungry Minded!