Has Mythos just broken the deal that kept the internet safe?

For nearly 20 years the deal has been simple: you click a link, arbitrary code runs on your device, and a stack of sandboxes keeps that code from doing anything nasty. Browser sandboxes for untrusted JavaScript, VM sandboxes for multi-tenant cloud, ad iframes so banner creatives can't take over your phone or laptop - the modern internet is built on the assumption that those sandboxes hold. Anthropic just shipped a research preview that generates working exploits for one of them 72.4% of the time, up from under 1% a few months ago. That deal might be breaking.

From what I've read Mythos is a very large model. Rumours have pointed to it being similar in size to the short lived (and very underwhelming) GPT4.5.

As such I'm with a lot of commentators in thinking that a primary reason this hasn't been rolled out further is compute. Anthropic is probably the most compute starved major AI lab right now and I strongly suspect they do not have the compute to roll this out even if they wanted more broadly.

From leaked pricing, it's expensive as well - at $125/MTok output (5x more than Opus, which is itself the most expensive model out there).

But this probably doesn't matter

One thing that has really been overlooked with all the focus on frontier scale models is how quickly improvements in the huge models are being achieved on far smaller models. I've spent a lot of time with Gemma 4 open weights model, and it is incredibly impressive for a model that is ~50x smaller than the frontier models.

So I have no doubt that whatever capabilities Mythos has will relatively quickly be available in smaller, and thus easier to serve, models.

And even if Mythos' huge size somehow is intrinsic to the abilities (I very much doubt this, given current progress in scaling smaller models) it has, it's only a matter of time before newer chips are able to serve it en masse. It's important to look to where the puck is going.

Sandboxing is at risk

As I've written before, LLMs in my opinion pose an extremely serious cybersecurity risk. Fundamentally we are seeing a radical change in how easy it is to find (and thus exploit) serious flaws and bugs in software for nefarious purposes.

... continue reading